Agentic Child Protection Investigation Assistant

Seized evidence in. Ranked, defensible leads out.

ACPIA is an air-gapped agentic system for authorised child-protection units. It correlates images, chat, audio, metadata and carrier records into a prioritised lead queue — and it gets there without forcing an analyst to look at abuse material first.

Never-view-by-defaultAir-gappedHash-chained audit
0
Innovation areas covered
0
Specialist agents
0 bytes
Evidence leaving the network
Abstract lattice of connected nodes, representing a fused evidence knowledge graph
Case CP-2291 · ingestrunning
Artifacts hashed
Metadata mapped
Graph fused
The problem

The evidence grew. The review process did not.

Four pressures compound in every child-protection case. ACPIA is designed against all four at once, rather than optimising one and ignoring the rest.

Volume

A single seizure can yield terabytes across phones, laptops, drives and cloud exports. Manual review does not scale to it.

Format sprawl

Chat databases, EXIF, call detail records, browser artifacts, voice notes, screenshots — each in its own dialect, none joined up.

Time

Safeguarding value decays fast. A lead found in week six is a different outcome from the same lead found on day one.

Human cost

Analysts absorb exposure that has documented psychological consequences. Any credible tool has to reduce it, not multiply it.

Dark server racks with patch cabling, representing large-scale evidence storage

The bottleneck was never storage.

It is the distance between a byte on a disk and a decision an investigator can defend in court. ACPIA is built to close exactly that distance.

Architecture

A planner, nine specialists, and a critic that says no

ACPIA is not one model answering questions. It is a supervised agent topology where every finding is re-checked against its source before an investigator sees it.

01

Evidence

Read-only mounts, SHA-256 addressing

02

Planner

Goal decomposed into a task graph

03

Specialists

Nine agents work in parallel

04

Critic

Unsupported leads are killed here

05

Investigator

Ranked queue with full provenance

Every arrow in this diagram writes a signed entry to the audit ledger.

Planner

Task plan

Decomposes the investigation goal into a task graph

Ingest

Artifact ledger

Mounts evidence read-only, hashes and indexes every artifact

Vision

Scene descriptors

Describes unknown media in text before anything is rendered

Audio

Transcripts

Transcribes, diarises and detects language and background cues

Text

Risk signals

Reads chat, mail and documents for intent and coercion signals

Metadata

Graph triples

Extracts and enriches EXIF, filesystem and carrier records

Correlation

Merged entities

Resolves entities and links sources across platforms

Timeline

Case timeline

Normalises clocks and merges events into one sequence

Critic

Verified leads

Re-checks every lead against source artifacts, kills the unsupported

Coverage

Twelve innovation areas, twelve working modules

Each module is a named service with its own inputs, outputs and confidence model — so a finding can always be traced back to the component that produced it.

Abstract technology imagery for the TRIAGE module
01TRIAGE

Content Analysis

Hash-first categorisation before a single pixel is rendered.

Known material is matched against agency hash sets and quarantined. Only unknown media reaches a model, and it is described in text before it is ever displayed.

Abstract technology imagery for the SENTINEL module
02SENTINEL

Threat Identification

Scores coercion and grooming patterns in conversation.

Linguistic escalation, isolation attempts and secrecy pressure are scored across chat, DM and SMS. It flags risk to a child — it never issues a verdict on legality.

Abstract technology imagery for the FUSION module
03FUSION

Source Correlation

Resolves one actor across many identities.

Handles, devices, IMEIs, numbers, wallets and IPs are reconciled into single entities with a confidence score and a full list of the evidence behind the merge.

Abstract technology imagery for the CONTEXT module
04CONTEXT

Contextual Extraction

Scene, object and audio cues from multimedia.

Room type, visible signage, uniforms, landmarks, on-frame OCR and speech transcripts combine into venue and geography hypotheses for safeguarding teams.

Abstract technology imagery for the PATTERN module
05PATTERN

Activity Pattern Analysis

Digital footprint rhythm and escalation curves.

Diurnal activity, platform-hopping, contact-frequency anomalies and escalation slope reveal behavioural signatures across a target's online life.

Abstract technology imagery for the TRACE module
06TRACE

Metadata Mapping

EXIF, filesystem and carrier metadata into a graph.

Every artifact contributes nodes and edges: device serials, GPS fixes, capture times, file lineage — enriched and mapped as a navigable relationship network.

Abstract technology imagery for the VERITAS module
07VERITAS

Synthetic Detection

Flags manipulated, synthetic or AI-generated media.

Diffusion and GAN artefact scoring, C2PA provenance inspection and sensor-noise consistency checks separate captured reality from generated content.

Abstract technology imagery for the CHRONO module
08CHRONO

Timeline Reconstruction

One timeline from many clocks.

Events from devices, servers and carriers are merged with timezone and clock-skew normalisation, so a sequence holds up when it is challenged.

Abstract technology imagery for the RECALL module
09RECALL

Intelligent Retrieval

Ask the case a question in plain language.

Hybrid keyword and local-embedding search across every extracted text surface. Answers arrive with citations that jump straight to the source artifact.

Abstract technology imagery for the SCRIBE module
10SCRIBE

Automated Reporting

Draft case reports where every line is cited.

Generates the case narrative, exhibit list and electronic-record certificate skeleton. Any sentence without a source artifact is refused, not guessed.

Abstract technology imagery for the PRIORITY module
11PRIORITY

Risk Assessment

Ranks leads by child-at-risk-now signals.

Recency, geographic proximity, contact-in-progress and escalation velocity decide the queue order — so the most urgent safeguarding lead is never buried at page nine.

Abstract technology imagery for the NEXUS module
12NEXUS

Intelligence Fusion

Finds the network no single artifact reveals.

Community detection and path analysis over the fused knowledge graph expose distribution rings and shared infrastructure across otherwise unlinked cases.

The console

Every lead arrives with its reasons attached

An investigator should never have to trust a score. Selecting any lead opens the exact artifacts behind it, the transform applied, and the confidence — ready for cross-examination.

ACPIA · Case CP-2291 · offline
No network egress

Why this lead

Active contact attempt from resolved identity KILO-3

  • Three handles resolved to one device (IMEI match, 0.91 confidence)
  • Message cadence increased 6× in the last 72 hours
  • Coercion-pattern score above threshold on 11 messages
  • Last activity 4 hours before seizure — contact likely ongoing

Recommended next step

Route to safeguarding desk for immediate welfare check

ACPIA proposes priority, never a determination. An authorised officer confirms or rejects — and that decision is logged.

Illustrative interface using synthetic case data. No real case material, victim data, or seized evidence appears anywhere on this site.

Safeguards

The constraints are the product

A tool that cannot be trusted in a courtroom, or that harms the analysts using it, has no operational value however accurate its models are.

Never-view-by-default

Known material is matched by hash and quarantined without display. Unknown media surfaces first as text, then blurred, and only at full fidelity under a logged two-person unlock.

Air-gapped by design

Open-weight models run on-premise on a single GPU workstation. No evidence byte ever leaves the agency network, and no third-party API is in the path.

Hash-chained audit ledger

Every agent action records its input hash, model version, output and operator into an append-only chain. Tampering breaks the chain and is provable.

Provenance on every claim

Each lead carries a ‘why this lead’ panel listing the exact artifacts behind it. Nothing is asserted to an investigator without a citation.

Human decides, always

ACPIA outputs priority and provenance — never a legal determination. No autonomous action is taken against any person under any circumstance.

Analyst wellbeing

Exposure time is tracked per session, audio defaults to off, and greyscale review is the default surface. Secondary trauma is treated as an operational risk.

An engineer working beside on-premise server infrastructure

One box. Inside your network.

Open-weight models on a single GPU workstation, installed from offline media. Nothing to send anywhere, nothing to leak.

Hard lines

  • No real abuse material is used, obtained or generated — in development, testing or demonstration.
  • ACPIA never generates, reconstructs, enhances or upscales abuse imagery.
  • Age estimation is a triage signal only, never a legal determination.
  • No autonomous action is taken against any person: no takedowns, no notifications, no arrests.
  • Deployment is limited to authorised agencies, scoped per warrant, with role-based access.
  • Every model output is rebuttable, logged, and attributable to a named operator.

Designed against the obligations agencies already operate under — POCSO 2012, IT Act s.67B, the electronic-record requirements of the BSA 2023, and DPDP 2023 data-minimisation. Section references should be re-verified against the current statute before any operational deployment.

Stack & roadmap

Built from parts that can run with the cable unplugged

Every component was chosen so the whole system can be installed offline, inspected by the agency, and defended line by line.

Agentic core

  • Planner–executor–critic loop
  • Tool-calling LLM runtime
  • Deterministic task graph
  • Per-case sandboxing

Perception

  • Open-weight VLM captioning
  • Local ASR + diarisation
  • OCR on frames & documents
  • PRNU / C2PA provenance checks

Knowledge

  • Property graph store
  • Local embedding index
  • Hybrid BM25 + vector recall
  • Entity resolution service

Forensics

  • Read-only evidence mounts
  • SHA-256 artifact addressing
  • Perceptual hash sets
  • Append-only audit chain

Interface

  • Lead queue & risk board
  • Interactive relationship graph
  • Timeline workbench
  • Cited report generator

Deployment

  • Single-GPU on-prem box
  • Air-gapped install media
  • Role-based access control
  • Per-case warrant scoping
Phase 0Now

Working spine: ingest, hash triage, metadata graph, timeline, cited retrieval, report generation.

Phase 1Next 8 weeks

Audio intelligence, synthetic-media scoring, and validation against an agency-held benchmark set.

Phase 2Pilot

Deployment with an authorised unit under an MoU, with independent review of the audit ledger design.

Phase 3Scale

Cross-case NEXUS fusion, multilingual coverage for Indian languages, and referral-format export.

Earth at night from orbit, city lights forming a network

A lead found on day one is a different outcome to the same lead found in week six.

ACPIA is built for authorised child-protection units, in partnership with the people who will have to defend its output. We are looking for pilot partners and domain reviewers.