# Rigr AI — Full Public Content for AI Agents > Secure AI triage for high-stakes evidence and safety workflows: media enrichment and triage for law enforcement and regulated investigations, plus product-grade apparent-age estimation for Trust & Safety teams. On-premise and air-gapped AI for video, image, audio, and text evidence. This file consolidates the public-facing content of rigr.ai into a single Markdown document for retrieval by AI agents and search systems. It reflects the English-language version of the site. Other languages (fr, de, es, pt, it) are available at `/{lang}/...`. All URLs and slugs are in English regardless of language. Gated content (admin dashboard, vetted-access login/request flows, presentation decks, and protected annex APIs) is **not** included here and should not be crawled — see https://rigr.ai/robots.txt. --- ## Home **Reduce digital evidence backlogs without compromising information control.** Rigr AI helps police and regulated investigative teams triage extracted evidence from phones, laptops, chat logs, images, video and audio at scale — across on-premise, air-gapped, private sovereign AI, secure private cloud and government cloud deployments. Digital evidence now dominates modern investigations. Reviewing it manually is slow, expensive, and often incompatible with operational timelines. Rigr AI provides a single platform that ingests raw media and converts it into structured, searchable, and reviewable evidence — without compromising security or evidential integrity. Rigr AI also serves investigators and intelligence analysts who need to estimate where an image was taken, Trust & Safety teams directly, with product-grade apparent-age estimation for content review routing, and US criminal-defense practices reviewing audiovisual discovery on hardware they control — see the Solution sections below (https://rigr.ai/solutions/digital-evidence-backlog, https://rigr.ai/solutions/criminal-defense-video-evidence, https://rigr.ai/solutions/trust-safety-age-estimation and https://rigr.ai/solutions/image-geolocation-investigations). **VST Teams (release v1.11.57) is a collaborative investigation platform.** Multiple analysts work the same case concurrently, with attributed bookmarks, annotations, and folder-level triage. It supports: - Rapid navigation of long video files with hover preview for triage - Accurate transcription, speaker identification, and translation - Face detection and extraction across video and still images - Text extraction from on-screen content (documents, signs, subtitles, overlays) - Automatic extraction of people, organisations, locations, and dates (Named Entity Recognition) - Integrated enrichment modules for age, location, and summarisation Built for regulated environments: on-premise, air-gapped, or cloud deployment; fully containerised; no external connectivity required in air-gapped mode; customer retains full control of data and outputs. Supports GDPR, the Law Enforcement Directive (EU 2016/680), and the EU AI Act (2021/0106). Rigr AI systems are used by national and supranational law-enforcement agencies in time-critical investigations, including large media seizures and cross-border cases. URL: https://rigr.ai/ --- ## Platform — VST Teams (v1.11.57) Investigations increasingly depend on large volumes of digital media. Manual review does not scale, and a single analyst cannot triage a seizure alone. VST Teams is a collaborative investigation platform. Multiple analysts work the same case concurrently, with named sessions, attributed bookmarks, and aggregated triage across folders. Deployed on-premise, air-gapped, or in your agency's cloud. Already in operational use. **What's new since v1.6:** - Multi-view incident review: every camera on one incident — body-worn, dashcam, CCTV, a bystander's phone, from any vendor — aligned on a single incident clock from their shared audio and played together. Spotlight layout with a thumbnail strip, frame-by-frame and slow-motion scrubbing, zoom and pan, a synchronised merged transcript that is reviewer-editable (rename a diarised speaker, toggle Original/Translated per line, correct a camera's source language and re-transcribe it), marked moments with labels and notes on the shared clock, and a composite MP4 export: every placed camera in one grid on one clock, with burnt-in camera labels and the transcript as subtitles. Incidents live inside a case. Shipped v1.11.57. - Known-material hash matching: import shared law-enforcement hash lists (INTERPOL, peer-agency, local) and match every file before ML classification runs; a listed item is resolved by the match rather than by model opinion. - A dedicated classification review workflow: a filterable, sortable triage queue with grid and focus modes, explicit verdict badges, and an end-of-queue session summary. - Documents as evidence: emails (.eml), Word documents (.docx) and plain-text files (.txt) are ingested, their contents extracted, made searchable and run through entity recognition; document language is detected from the text. These three are the supported document formats. - Search reaches what the AI found: a word spoken in a video, shown on screen, or a named face now finds the file, where previously only filenames and folder names matched. **Shipped in v1.6 (April 2026):** - Multi-user workflows with named sessions, per-user bookmarks and annotations, and licence enforcement at login - Automatic entity extraction — Named Entity Recognition for people, organisations, locations, and dates; entities can be merged, split, relabelled, and exported - Still images processed alongside video in a unified gallery — no separate workflow - Folder-level triage panel with aggregated file counts, detected faces, text regions, and processing status - Hover preview for rapid file triage - Chunked uploads with automatic retry **What the platform does.** It ingests raw media and produces structured outputs that investigators can search, review, and reference. It supports: - Video evidence from seized devices, body-worn cameras, and online sources - Still images processed alongside video in a unified gallery - Audio and text evidence, including multilingual material - Folder-level aggregation for triage across large seizures The output is an investigative workspace, not a black-box result. **Core capabilities:** - **Video navigation and rapid triage** — long videos are segmented into navigable thumbnails; hover preview shows a silent thumbnail before opening any file. - **Transcription and speaker identification** — accurate transcripts from audio and video with speakers labelled separately; automatic language detection and translation. - **Face detection and extraction** — faces detected, grouped, and presented in a structured gallery across video and still images. - **Text extraction (OCR)** — on-screen text made searchable: documents, signs, phone screens, subtitles, usernames, overlays. - **Automatic entity extraction (NER)** — people, organisations, locations, and dates extracted from transcripts and on-screen text. - **Multi-user workflows** — named user accounts with individual sessions; bookmarks and annotations attributed per analyst; licence enforcement at login. - **Search, analysis, and reporting** — full-text search filtered by face, entity, keyword, or time; exportable structured reports. **Enrichment modules** integrate directly into the platform, and are also available as standalone APIs: - Age Estimation — highlighting potentially sensitive content involving minors - Image geolocation (Morrigan) — estimating where a photo or video was taken, from the scene itself - Text & Audio Summarisation — rapid triage of transcripts and chat logs **Deployment:** on-premise, air-gapped, or cloud; fully containerised — runs anywhere containers run; no proprietary infrastructure. Capability brief (police edition), PDF: https://rigr.ai/resources/vst-teams-capability-brief-police.pdf Earlier v1.6 printed feature list, kept at its original URL: https://rigr.ai/resources/vst-teams-v1.6.pdf ### Capability Brief (Police Edition) — full text summary A PDF cannot be read by an agent, so the brief's substance is reproduced here. **Positioning.** Reduce digital evidence backlogs without compromising information control — turning large volumes of video, images and text into usable evidence, on-premise or air-gapped. Status: in operational use. Function: enrichment and triage. Deployment: on-premise, air-gapped, cloud. **Deployment for operational teams.** Installed by the unit's own technical officer with a Rigr engineer hands-on throughout, by remote session where that helps. Investigators just open VST; no Docker, no terminal. **Core capabilities.** Video navigation and triage (skim a three-hour video via segmented, hover-preview thumbnails linked to timestamps); transcription with speaker separation and automatic language identification (in incident review, a reviewer can correct a camera's source language and run transcription again for that feed); face detection and extraction, grouped by person; text extraction from documents, signs, phone screens, subtitles and overlays; entity extraction of people, organisations, locations and dates, with merge/split/relabel/export; multi-user workflows with named sessions, per-analyst bookmarks and annotations, and folder-level triage; full-text search filtered by face, entity, keyword or time, with exportable structured reports. Multi-view incident review — every camera on one incident aligned on a single timeline from shared audio — is released, shipped in v1.11.57. **Enrichment modules**, available inside the platform or as standalone APIs: age estimation, image geolocation (Morrigan), and text and audio summarisation with translation. **Deployment and information control.** On-premise, air-gapped, private sovereign AI, secure private cloud or government cloud. Ingestion by watched folders or API, integrating with existing evidence platforms. Self-contained on customer hardware with no proprietary infrastructure to procure, and no external data dependency in air-gapped mode. Data and outputs remain under customer control. Seat-based floating licences enforced at login. Designed for GDPR and the Law Enforcement Directive (EU 2016/680), built to align with the EU AI Act. **Where Rigr fits.** Existing forensic tools acquire, extract or export device data — Rigr does not replace acquisition, unlocking or forensic collection. Rigr ingests and structures that material for search, review and prioritisation. Investigators review priority material, validate outputs and build evidential narratives. Findings and reports export into downstream case and evidence workflows. **Media classification (available now).** An entire seizure is read automatically, ranked by severity, with the material that matters brought to the top of the queue. Severity categories are 0 (none), 2 (exploitative), 3 (posing), 4 (non-penetrative) and 5 (penetrative); level 1, adult-benign, is a reviewer verdict rather than a classifier output. Content with no sexual content detected — family photographs, documents, scenery — is scored zero, which orders the review queue rather than deciding relevance: such an item can still matter to a case, and can be filtered back in. Content is blurred by default, with optional greyscale, and hold-to-reveal and per-item blur a single key away, because reducing the volume and duration of exposure is treated as a first-order design requirement. The model proposes a classification; a trained reviewer confirms, overrides or disputes it, with a full audit trail. **Redaction for disclosure (preview).** Identities of victims, informants and uninvolved third parties must be removed before evidence is shared. VST surfaces every candidate and states plainly that it may not have caught them all; a reviewer redacts, pseudonymises or keeps each one; informant-identity candidates are set apart and default to full removal; one decision applies across every layer with a full audit. This is a preview of an upcoming release, not a shipped capability. **Sector context.** The House of Lords Science and Technology Committee has highlighted that digital forensics backlogs continue to undermine timely justice, and pointed to trusted AI and digital tools as part of the response. The output is an investigative workspace — searchable, reviewable and exportable — not a black-box result; investigative and evidential judgement stays with the customer's team. URL: https://rigr.ai/platform --- ## Solution — Digital Evidence Backlogs Buyer-led page for police, digital forensics leads, and regulated investigative teams. Digital evidence — phones, laptops, chat logs, images, video and audio — now produces more material than human-only teams can manually review on operational timelines. That pressure shows up as safeguarding decisions, charging timelines, bail decisions and case progression all waiting on a review queue, not as a single team falling behind. The House of Lords Science and Technology Committee has highlighted that digital forensics backlogs continue to undermine timely justice, and pointed to trusted AI and digital tools as part of the response. Rigr AI sits after acquisition/export and before deep manual review: a secure AI triage layer that structures extracted media, chats, transcripts, images, video and audio for search, review and prioritisation, directing scarce specialist time to what's most likely to matter. It does not replace acquisition, unlocking, or forensic collection tools, and it does not replace human investigative or evidential judgement — final decisions remain with authorised humans. URL: https://rigr.ai/solutions/digital-evidence-backlog --- ## Solution — Criminal Defense Video Evidence Buyer-led page for US private criminal-defense attorneys, assigned counsel and public-defender teams. A routine criminal matter now arrives with hours of audiovisual discovery: body-worn camera, dashcam, police interviews and interrogations, jail calls, 911 and dispatch audio, CCTV and phone video, often multilingual. Commercial transcription and translation are billed by the minute, and watching everything in real time consumes the hours available for the case. VST transcribes, translates and summarises that material on an NVIDIA workstation inside the firm, producing timestamped searchable transcripts with speaker separation, a working translation alongside the original, and summaries that tell the attorney which recordings are worth their time. A summary is a triage signal, not a finding: the relevant transcript and recording should be checked before relying on it. Nothing has to be uploaded to a shared public AI service. Commercially this is a flat annual small-office licence with no per-minute transcription charge, running on a single processing workstation rather than an enterprise server. VST produces an AI-assisted working transcript for review and preparation; it does not replace a certified transcript where one is required, and machine translation does not replace certified translation for formal evidential use. Synchronised multi-camera playback on a single incident clock is released; alignment depends on usable shared audio, and recordings VST cannot place confidently are flagged for manual placement. URL: https://rigr.ai/solutions/criminal-defense-video-evidence --- ## Solution — Trust & Safety Image Moderation Buyer-led page for Trust & Safety, platform policy enforcement, and AI safety teams. Media Examiner detects nudity, exposed intimate anatomy and sexual activity while estimating apparent age and developmental stage — all in one image-analysis API call. There is no chaining a generic adult-content classifier to a separate age service and no submitting the image twice. Generic nudity detection is a commodity; the differentiator is richer child-safety and policy signal from a single analysis: what triggered the result, where it is in the frame, the apparent developmental stage of the subject, and where a face is visible, the estimated age and its uncertainty. One call returns: an image-level severity classification (0 no sexual content, 2 exploitative or suggestive, 3 overt sexualised posing, 4 non-penetrative sexual activity, 5 penetrative sexual activity); per-object detections with confidence scores and bounding boxes; apparent developmental stage (infant, toddler, prepubescent, pubescent, adult) from the face or, where the face is hidden, from the body; apparent age with calibrated uncertainty on face detections; and contextual flags (self-generated, sadomasochism, CG elements, bodily fluids). Available as a hosted API, a private endpoint, or a container the customer runs, including air-gapped. Recommended upload workflow: auto-handle clear violations; escalate anything where the subject may not be an adult; route low-confidence and borderline cases to human review; retain the per-result detections, scores and boxes as the audit trail for why an action was taken; and tune thresholds to the platform's own risk tolerance. Rigr deliberately does not publish a single universal threshold, because a vendor offering one is making the platform's policy decision for it — the configurable review band is the product. Apparent-age and developmental-stage outputs are visual estimates. They do not verify identity and do not establish a person's legal age. They are a signal for routing and prioritisation, not a determination; high-risk and borderline decisions should include appropriate human review, and confirmed child sexual abuse material must be handled under the platform's legal reporting obligations rather than simply removed. Where this sits in age assurance: age assurance is the umbrella term for establishing how old someone is, and its methods are not interchangeable. Age verification checks an identity document or an authoritative record to establish a stated age — Rigr does not provide it. Facial age estimation infers an apparent age from an image with a confidence range — this is what Rigr builds. Age inference from behaviour or account history, and self-declaration, are also outside what Rigr provides. Rigr is therefore a facial age estimation component, not an end-to-end age assurance stack: a relying party integrates it as one signal in a workflow they own, usually alongside a document-verification route for cases that need a proven age, and always with a human path for borderline ones. An obligation to verify a specific person's legal age is not discharged by estimation alone. Estimation applies where no document exists and no account is attached — user uploads, AI-generated characters, and imagery already in a moderation queue, where it is the only method that applies at all. Age assurance regimes generally expect a relying party to manage uncertainty rather than act on a single hard cut-off, which is what the calibrated uncertainty is for: a configurable buffer zone around the threshold, with everything inside it routed to a human. The Australian Age Assurance Technology Trial specifically noted this design. Rigr publishes no universal threshold, because the same number should not trigger the same action on a children's education service and an adult platform. Trust & Safety teams do not just need a binary "adult"/"minor" label — they need to know how uncertain the system is, how to route borderline content, and how to defend the threshold used for human review. AI-generated and manipulated imagery has made age-ambiguous content review harder to staff and easy to get wrong in either direction; generic prompting against a general-purpose vision-language model is not a calibrated, defensible safety control. Rigr returns an **apparent-age** point estimate plus **calibrated uncertainty** and face-level metadata, so platforms can route high-risk and borderline content to the right review queue rather than a single undifferentiated pool, with human-in-the-loop escalation for borderline and low-confidence cases. A private benchmark pilot is available for teams currently relying on generic VLM prompts, off-the-shelf classifiers, or manual-only queues. URL: https://rigr.ai/solutions/trust-safety-age-estimation --- ## Solution — Image Geolocation for Investigations For victim identification, OSINT and digital-investigation units, and intelligence and defence analysts — anyone whose question about an image is *where was this taken?* Morrigan estimates likely capture locations for images and video, indoors as well as outdoors, from the whole scene (architecture, materials, signage and script, vegetation, terrain, light, interior furnishings, clothing) rather than waiting for a landmark. It returns several candidate locations ranked with confidence scores; a spread across regions signals that the image carries little location evidence. The investigator UI adds joint estimation across several images believed to share a location (materially more accurate than any single image), video frame by frame, and a crop editor. Runs inside VST Teams across a whole case, as a standalone REST API, in the browser-based investigator UI (English interface), or on-premise and air-gapped as a container. Results are estimates for an analyst to confirm against other sources — not findings about where an image was taken. The public demo reads one still image at a time (three runs a day; submissions kept up to 7 days for trust and safety review), and offers a self-serve trial key for the full investigator tool. URL: https://rigr.ai/solutions/image-geolocation-investigations --- ## Capability — Age Estimation Estimate apparent age from a face, with uncertainty alongside the result. Use estimates to prioritise investigative review or route age-ambiguous content for trust and safety teams. A visual estimate does not establish a person's legal age. Try an image in the public demo: https://rigr.ai/capabilities/age-estimation/#try-age Read the facial Age Estimation API reference: https://rigr.ai/docs/age-estimation-api/ For age estimation in the same call as media classification: https://rigr.ai/docs/media-examiner-api/ **Evaluation:** Rigr's benchmark reports 1.64 years overall MAE and 0.87 years for ages 0–18. Error varies by age group, image quality and population; these are dataset results, not a guarantee for every upload. **Operational use within VST Teams:** highlight potentially sensitive content involving minors; prioritise review across large media datasets; support evidential assessment without replacing human judgement. Also available as a standalone API or lightweight application. **On-premise deployment:** fully containerised, including air-gapped operation. Processing stays within the customer's infrastructure; no images are sent to Rigr AI. The public demo is separate: submitted images are kept for up to seven days for trust and safety review, then deleted. Policy: https://rigr.ai/privacy/age-demo/ URL: https://rigr.ai/capabilities/age-estimation/ --- ## Capability — Image Geolocation (Morrigan) Visual geolocation from images and video, indoors and outdoors, ranked with confidence scores. Rigr AI's image geolocation capability (Location Estimation, product name Morrigan) analyses visual cues in images and video to estimate likely capture locations, both indoors and outdoors. Results are probabilistic and ranked with confidence scores, supporting lead generation rather than false certainty. **Operational use within VST Teams:** support jurisdictional assessment; generate investigative leads from visual evidence; accelerate cross-border and time-sensitive investigations. Also available as a standalone API and through Morrigan's dedicated investigator UI (https://demo.morrigan.rigr.ai/ — browsable publicly; running analysis requires a trial API key, issued via the capability page below). The investigator UI supports video as well as still images, frame cropping, and joint estimation across multiple media items. The capability page hosts a free cut-down public demo (still images, daily quota). **Deployment:** fully containerised, on-premise and air-gapped operation, no external connectivity required, outputs remain under customer control. URL: https://rigr.ai/capabilities/location-estimation --- ## Capability — CSAM Classification (Media Classification) AI-powered detection and classification of sexual content in images, purpose-built for CSAM investigations and content moderation at scale. Identifying and classifying sexual content in seized media is central to child exploitation investigations. Manual review is slow, traumatic for investigators, and inconsistent at scale. Rigr AI's Media Classification capability uses deep-learning object detection to identify sexual content, body parts, activities, and contextual indicators in images — then maps those detections to a structured severity classification aligned with established review frameworks. **Operational use within VST Teams:** triage large volumes of seized media by severity; identify specific sexual activities, body parts, and contextual indicators; flag self-generated content, CSAM network logos, and other evidentiary markers; prioritise review queues so investigators focus on the most serious material first. Also available as a standalone API or containerised application that can be deployed in an air-gapped environment. **Deployment:** fully containerised, on-premise and air-gapped operation, no data retention, customer retains full control of inputs and outputs. **What makes it different:** severity is not a binary "CSAM or not" — it is graded on the specific act (from posing through to penetrative) plus aggravating flags (a weapon, violence, animals), so teams can apply their own jurisdiction's threshold. The hardest call — whether the subject is a child — is estimated from the face (the strongest signal) and from the body when the face is hidden. Known-material hash matches are shown with provenance and kept separate from the model's opinion. Everything is blurred by default and every classification is confirmed by a person. Beyond triage, an early read of the type and severity of a collection supports interview preparation and risk/safeguarding assessments (e.g. a collection focused on young boys is exactly what a risk assessor or social worker needs to know). **Interactive tour (canned, safe to explore):** https://rigr.ai/tours/classification-review/ — a click-through walkthrough of the real VST review screens: triaging a whole case, filtering to what needs a person, deciding on severity, and known-material matches. No backend, nothing is processed. Localised into all ten site languages at /{lang}/tours/classification-review/. Tour index: https://rigr.ai/tours/ **Video walkthrough (3 min 8 s):** https://rigr.ai/watch/media-classification/ — Edward Dixon narrates media classification in VST Teams on a development build, using a test set of adult images that stay blurred throughout. Part 1, severity scales: a jurisdiction profile maps the classifier's detections (body parts, nudity, sexual acts, fluids) onto the team's own scale, starting from a standard scheme if wanted; an administrator can change detection thresholds where a label fires too readily or too rarely. The Poland profile shown is an example for that test installation. Part 2, review: the model proposes a severity for each file and shows which detections produced it; reviewers filter by certainty, confirm in bulk, open a focused view to see why a file was graded as it was, reject a single detection or override the grade. Part 3, report: export for a case or a folder, in four formats including Excel, listing the proposed and effective severity, the reasons and who reviewed each file. English narration. The page, captions and transcript are localised into all ten site languages at /{lang}/watch/media-classification/. URL: https://rigr.ai/capabilities/media-classification --- ## Capability — Text & Audio Summarisation Secure summarisation and translation of transcripts, chat logs, and multilingual text evidence. Large volumes of text and audio evidence are now routine in modern investigations. Manual review does not scale. Rigr AI's Text & Audio Summarisation capability securely condenses transcripts, chat logs, and multilingual material into concise summaries suitable for rapid triage and evidential preparation. **Operational use within VST Teams:** rapid triage of long transcripts and chat histories; multilingual translation and summarisation; identification of key statements and entities. Designed for sensitive and challenging content that general-purpose AI systems often reject. **Deployment:** fully containerised, on-premise and air-gapped deployment, no external data dependency, customer retains control of all data. URL: https://rigr.ai/capabilities/summarisation --- ## Project — ARICA Assessing Risk Indicators of Child Abuse, developed under EU funding. ARICA is a grant-funded project developing advanced AI capabilities to support law enforcement agencies investigating child sexual exploitation material (CSEM). Grant funding has come from the European Union, from EACTDA (European Anti-Cybercrime Tools Development Agency) and from End Violence, with UK Home Office funding. The project brings together technology providers, law enforcement agencies, and academic partners to deliver tools that accelerate risk analysis and triage, improving targeting decisions. **Rigr AI's role:** contributes core AI capabilities to ARICA, including enrichment of text, images, and video. **Impact:** ARICA directly supports investigators working on some of the most challenging cases in law enforcement. By automating initial triage and providing reliable AI-assisted assessments, the project aims to surface the most actionable data as soon as possible and accelerate identification and safeguarding of victims. Public project site: https://aricaproject.eu A protected Law Enforcement Technical Annex is available to vetted police users via https://rigr.ai/access/request. URL: https://rigr.ai/projects/arica --- ## Project — BadB GNSS-denied navigation and positioning for defence applications, developed under EU funding. BadB is an EU-funded defence research project focused on uncrewed-vehicle autonomy in contested areas. The project addresses scenarios where GNSS signals are unavailable, jammed, or spoofed, and other radio communication may be difficult or dangerous. Modern operations — both military and civilian — depend heavily on accurate positioning. When satellite signals are compromised, alternative approaches are needed to maintain operational effectiveness. **Rigr AI's role:** contributes to the software, simulation, and machine learning aspects of BadB. **Applications:** - Defence operations — maintaining navigation capability in contested electromagnetic environments, enabling ISR and other mission types - Border security — resilient drones to enhance border surveillance and control - Emergency services — enabling drone use in civil spaces where anti-drone systems may also be in use (policing, prison service, etc.) Public project site: https://badbproject.eu A protected Defence & Intelligence Annex is available to vetted defence users via https://rigr.ai/access/request. URL: https://rigr.ai/projects/badb --- ## Articles Edward Dixon's occasional writing on AI, investigation tooling, defence research, and the systems Rigr builds. Long-form, low cadence (roughly monthly to bi-monthly), not promotional. Articles index at https://rigr.ai/articles/. RSS feed at https://rigr.ai/articles/rss.xml. This section is intentionally simple by design: markdown source files, version-controlled via git, rendered through the same Astro site, with JSON-LD `BlogPosting` per article and an RSS feed for syndication. Per-article OG images and meta tags follow the same conventions as the rest of the site. When new articles are published, brief summaries may be added here for AI retrieval. The canonical full content lives at the article URLs themselves. --- ## Interactive demos ### Air Defence Allocation Sim Browser-based, deterministic simulation exploring the trade-off between fixed point air defence, mobile combat air patrol (CAP), sensor / early-warning depth, and saturation attacks. Built by Rigr to make the trade-offs of an integrated air defence concrete and visible. The thesis being explored: fixed defences are valuable when target value is concentrated, but as cheap long-range precision strike makes more targets worth attacking, defence must become more mobile, networked, and able to concentrate against the main attack. Defender controls: total defence budget, mobile-vs-fixed share, warning range multiplier. Attacker controls: weapons per wave, wave count, concentration (one capital vs spread across many towns), doctrine. Output metrics include preserved target value, leakers, intercept geometry, mobile fighter magazine utilisation, mean warning time, and concentration efficiency. The sim is abstract — no real systems, weapons, performance figures, or countries are modelled. Settlement names, country shape, and road network are procedurally generated per seed. Visual model: light-cartographic ("map") view by default, with a dark "tactical" theme toggle. The country is rendered with land/sea coastline, OSM-style road hierarchy (motorway and secondary), built-up area shading, and procedurally-named cities including a capital. Live engagement shows strike weapons inbound, sensor coverage rings, fixed defence batteries firing missiles, mobile CAP fighters flying to screen positions and engaging, and intercepts and impacts as ring animations. Includes scenario presets (Standard saturation, Surprise volley, Capital strike, Distributed precision, Forward-warning success, Saturation overload), a timeline scrubber with pause and seek, a jump-to-first-contact control, and PNG export at 2× resolution for sharing or embedding in articles. Every run generates a written "Reading this run" analysis and a same-seed counterfactual: the identical attack replayed with the mobile/fixed allocation flipped, so the value of the chosen posture is shown against the same dice. URL: https://rigr.ai/sim/ ### Convoy Protection Overwatch Sim Browser-based, deterministic simulation of defending a small armoured patrol (2–8 vehicles) against drone ambush launched from close terrain. Built by Rigr to make the counter-UAS protection problem for ground forces concrete: in close country the treeline — not the interceptor — sets the warning time. The thesis being explored: organic hard-kill on the vehicles is necessary but not sufficient. The sensor horizon set by terrain standoff is usually shorter than a realistic detect–decide–engage chain; saturation defeats Pk because every drone beyond per-window shot capacity is a guaranteed leaker; and the engagement that decides the fight has to happen above the trees — sensor-only overwatch pre-loads the reaction timer, while armed overwatch changes the geometry itself by engaging over the canopy. Terrain line-of-sight is 2.5D: polygonal forest blocks carry a uniform canopy height (user-tunable, 10–30 m), and every sight line is ray-checked against entity heights — 3 m hull sensors, a 25 m telescopic mast head, a 400 m overwatch platform, and a drone altitude profile (a launch climb above the canopy that elevated sensors can catch, a masked run beneath the drone's own treeline, low cruise in the open, and crests over any other woodland crossed). The resulting **dead ground** — terrain where a cruising drone is invisible to every sensor in the column — is shaded live on the map, with a separate tint for the (near-empty) area hidden from the overwatch platform. Hull sensors treat every treeline as a wall; a mast above the canopy turns the wall into a finite skirt (multiplier (mast−cruise)/(mast−canopy)); overwatch at altitude has effectively no dead ground beyond the under-canopy interiors of the woods themselves. Defender controls: vehicle count, spacing, and speed; per-vehicle hard-kill engagement range, reaction time, single-shot Pk, magazine, fire channels, and open-ground sensor range; and a top-cover layer (none, telescopic sensor mast, sensor-only UAV overwatch, armed overwatch). Attacker controls: drone count and speed, number of ambush events, single-axis vs two-axis pincer launch, and concentrate-vs-spread targeting. Terrain controls: forest standoff from the road, forest density, and canopy height. Output metrics include vehicles lost, leakers, kills by convoy vs overwatch, mean and worst-case warning time, saturation events (tracked drones in range with no free fire channel anywhere in the column), shots fired, realised Pk, and rounds remaining. Every run is simulated headlessly to completion before the first frame is drawn, then replayed deterministically: timeline scrubber, jump-to-first-contact, playback speeds, and PNG export. Each run is automatically paired with a same-seed armed-overwatch counterfactual — the identical attack, same dice, with an armed air layer — and a written analysis of the sensor-horizon arithmetic for the chosen parameters. Visual model matches the Air Defence Allocation Sim: light-cartographic "map" view by default with a dark "tactical" theme toggle; winding road, procedurally generated forest blocks and ambush sites, engagement and sensor rings, the dead-ground overlay (toggleable), tracer lines, kill rings, and impact animations. Scenario presets: Roadside ambush, Two-axis pincer, Saturation swarm, Close country, Armed overwatch, and 75-drone pack (a commercial swarm pack; the drone slider runs to 80). The sim is abstract — no real-world systems, weapons, or performance figures are modelled. EW, acoustics, and soft-kill are deliberately out of scope. URL: https://rigr.ai/sim/convoy/ --- ## Contact For briefings, demonstrations, or technical discussions: https://rigr.ai/contact --- ## Get a free API key Self-serve signup for a free Age Estimation API key — a lighter-weight path than VST Teams for individual developers and investigators who want to call the API directly. Sign up with a work email and the key arrives by email, ready to paste into the Settings panel at https://demo.age.rigr.ai/ or call the API directly. URL: https://rigr.ai/get-api-key --- ## Independent evaluation — Australian Age Assurance Technology Trial Rigr AI participated in the Australian Age Assurance Technology Trial (2025), an independent evaluation commissioned by the Australian Government and conducted by the Age Check Certification Scheme. Participants were tested against a dataset built by the trial rather than against their own benchmarks. The trial's assessment of Rigr AI found strong technical readiness, privacy alignment and usability, with reliable threshold-based age decisions, fast on-device processing and low false-positive rates in real-world scenarios. It noted the privacy-centric architecture — no personal data stored, biometric inputs processed and deleted locally in milliseconds after estimation — and configurable buffer zones allowing a relying party to manage uncertainty rather than depend on a single hard threshold. Primary sources: https://ageassurance.com.au/v/rgr/ · https://ageassurance.com.au/wp-content/uploads/2025/08/IndividualTestReport-RigrAI_AE.pdf Mean-absolute-error figures are not comparable between evaluations run on different datasets. Rigr's own benchmark figures (1.64 years overall, 0.87 for minors, 68.9% within one year) are measured on Rigr's dataset and should not be read against results from the trial's dataset or any other. --- ## Age Estimation API reference Full REST reference for the facial age estimation API, previously published only as a PDF. The API accepts base64-encoded images and returns, for every detected face, an estimated age in years, a calibrated uncertainty value in years, a bounding box, and a detection confidence score. The current model (v3) achieves a mean absolute error of approximately 1.6 years on Rigr's benchmark dataset. Endpoints: GET /healthz (unauthenticated health check), POST /api/image (submit images for estimation, with optional detection_threshold and max_dets_per_image), GET /api/info (model version, batch and image limits, detector/estimator models and devices), GET /api/stats (throughput statistics). All endpoints except /healthz require an API key in the X-API-KEY header. Image limits: PNG, JPEG, JPG, BMP, GIF or WebP, base64-encoded, 50 MB maximum per image, between 24x24 and 6000x6000 pixels. Errors are structured JSON with type, message and detail, returned either at the top level (request could not be processed) or per image (one image failed while the rest succeeded). A truncated image is still processed, with the error field flagging it. Deployment: hosted cloud at https://api.age.rigr.ai for evaluation, or self-hosted Docker images for CPU, GPU (CUDA) and ONNX-only CPU inference, all exposing the same API on port 8000 and supporting air-gapped operation. URL: https://rigr.ai/docs/age-estimation-api --- ## Media Examiner API reference Media Examiner detects nudity, exposed intimate anatomy and sexual activity while estimating apparent age and developmental stage — all from one image-analysis API call. There is no need to chain a generic adult-content classifier to a separate age service, and no need to submit the image twice. Base URL https://api.mes.rigr.ai. Authentication is an API key sent as either an X-API-KEY header or an Authorization: Bearer header; the two are equivalent. Images are sent as multipart/form-data. Endpoints: POST /classify (image-level severity classification with the detections that justify it — the integrated call), POST /predict (raw object detections with bounding boxes, no severity roll-up; /upload is a legacy alias), POST /detect-age (faces and their ages only, in a response schema matching the standalone age service so existing clients move across by changing the URL), POST /estimate-age (a single age estimate for a pre-cropped face). Request fields for /classify and /predict: file (required), model (default VisualyzeV2, also FullBodyLarge), score (detection threshold, default 0.25), iou (default 0.45), topk (default 100), estimate_age (default false — set true to run facial age estimation on detected faces in the same inference pass; age is opt-in and does not run unless requested). The /classify response carries: classification.key (machine-readable, stable), classification.display_name, classification.severity; flags (the union of contextual flags across detections); and detections, each with class_name, score, bbox, ucs_sexual_content (the abstract category, null for classes carrying no sexual-content meaning such as faces, clothing or life-stage classes), and ucs_flags. With estimate_age=true, face detections additionally carry age (apparent age in years), age_uncertainty (calibrated standard deviation in years), and original_class_name where the age model overrode the detector's bracket label. Severity scale: 0 rigr-none (no sexual content), 2 rigr-exploitative (nudity, exposed anatomy or sexualised context without explicit activity), 3 rigr-posing (overt sexualised posing), 4 rigr-non-penetrative (masturbation, licking, other non-penetrative contact), 5 rigr-penetrative (intercourse, oral sex, anal or vaginal penetration). The classifier never emits severity 1: level 1 exists in the wider review scale as an analyst verdict for explicit adult material judged benign on review, which no model assigns. This is a triage scale, not a statutory one — categories and thresholds differ by jurisdiction and should be mapped to local rules rather than treated as a universal decision boundary. Contextual flags are orthogonal to severity and describe the character of the material rather than how explicit it is. Four values are emitted: Self-Generated, Sadomasochism, CG Elements, Bodily Fluids. Apparent developmental stage brackets: Infant 0 to 0.9, Toddler 1 to 4.9, Prepubescent 5 to 12.9, Pubescent 13 to 17.9, Adult 18 and over. Bounding boxes on /classify and /predict are normalised x/y/w/h objects with a top-left origin — fractions of the frame, not pixels, and not a corner pair. /detect-age is the exception and returns pixel [x_min, y_min, x_max, y_max], because it mirrors the standalone age service's schema. Image limits: JPEG, PNG, GIF, WebP, BMP, TIFF, AVIF, HEIC and HEIF are accepted; animated PNG and GIF are analysed frame by frame with per-frame detection blocks; 50 MB maximum per upload, one file per request. Frames above roughly 24 megapixels are downscaled proportionally before inference rather than rejected. Rate limits: 120 requests per 60 seconds per API key by default. Every response carries x-rate-limit-remaining; a 429 includes retry_after_seconds. Recommended backoff is exponential, 1s, 2s, 4s, up to 30s. Errors: 400 empty or undecodable image, 401 missing_api_key or invalid_api_key, 413 upload over 50 MB, 415 unsupported image type, 429 rate_limited, 500 inference_unavailable (transient, retry), 503 age model not loaded on this deployment. Limitation: apparent-age and developmental-stage outputs are visual estimates. They do not verify identity and do not establish a person's legal age. High-risk and borderline decisions should include appropriate human review. The calibrated age_uncertainty is what makes routing possible — a tight interval well clear of a threshold can be decided automatically, a wide one straddling it should go to a human. Deployment: hosted at https://api.mes.rigr.ai, or as a container exposing the same REST API plus a gRPC interface for streaming uploads, supporting on-premise and air-gapped operation. Machine-readable contract: https://rigr.ai/openapi/media-examiner.yaml (OpenAPI 3.1). URL: https://rigr.ai/docs/media-examiner-api --- ## Site map (public) - https://rigr.ai/ - https://rigr.ai/platform - https://rigr.ai/solutions - https://rigr.ai/solutions/digital-evidence-backlog - https://rigr.ai/get-api-key - https://rigr.ai/docs/age-estimation-api - https://rigr.ai/solutions/criminal-defense-video-evidence - https://rigr.ai/solutions/trust-safety-age-estimation - https://rigr.ai/solutions/image-geolocation-investigations - https://rigr.ai/capabilities - https://rigr.ai/capabilities/age-estimation - https://rigr.ai/capabilities/location-estimation - https://rigr.ai/capabilities/media-classification - https://rigr.ai/capabilities/summarisation - https://rigr.ai/projects - https://rigr.ai/projects/arica - https://rigr.ai/projects/badb - https://rigr.ai/deployment - https://rigr.ai/why-rigr-ai - https://rigr.ai/use-cases - https://rigr.ai/contact - https://rigr.ai/articles/ - https://rigr.ai/sim/ - https://rigr.ai/sim/convoy/ Localised equivalents are at `/fr/`, `/de/`, `/es/`, `/pt/`, and `/it/` with English slugs (e.g. `/fr/platform`). Full XML sitemap: https://rigr.ai/sitemap-index.xml.