The Fact Pilot
AI-assisted factual investigation, built for insurance and legal work.
A multi-tenant SaaS that turns raw interview audio into court-ready statements, chronologies, and reports. Australian data residency and per-firm database isolation from day one.
Give factual investigators a single, secure workspace.
The context
An Australian factual-investigation firm working across insurance and legal matters needed to move off fragmented tooling. Interview recordings, transcripts, statements, and reports were spread across disconnected apps, with compliance and data-residency risk on every handoff.
The mandate
Ship a production SaaS that investigators trust: on-shore data, per-firm isolation, and AI that drafts structured witness statements and investigation reports without fabricating facts.
The bar
Build the pipeline investigators use every day: resilient uploads, diarised transcripts, AI extractions bound to a pre-agreed structure per participant type, and artefact exports that hold up in review.
Three months, production-ready.
What we did
- Multi-tenant architecture with database-per-firm isolation
- End-to-end interview pipeline: upload → transcribe → extract → draft → export
- Role-based workspaces for investigators, reviewers, and firm admins
- Artefact generation (PDF + DOCX) with firm branding
- Tenant-level usage controls and audit trails
- Australian-region deployment with MFA-protected access
Our process
Mapped the investigator workflow end-to-end, nailed down data-residency and isolation requirements, and designed the tenant model before writing production code.
Built the upload → transcription → AI extraction loop against a pre-agreed output shape per participant type (Claimant, Insured, Witness), so outputs stay structured and reviewable.
Shipped investigator, reviewer, and admin workspaces, plus PDF / DOCX statement and report generation with firm branding and usage caps.
Locked down TOTP MFA, per-tenant usage caps, and audit logging, then shipped the production deployment to the first firm.
Services covered
A stack chosen for trust, residency, and velocity.
- Next.js (App Router)
- React
- Tailwind
- Node.js
- Express
- Mongoose
- MongoDB
- Database-per-tenant
- Vertex AI
- Gemini 2.5 Flash
- australia-southeast1
- AWS Transcribe (diarization)
- S3 (ap-southeast-2)
- PDFKit
- docx
- html-to-docx
- TOTP MFA (speakeasy)
- AWS Elastic Beanstalk
Deployment pipeline
Stack summary
- Structured outputsEach AI call is bound to a pre-agreed output shape per participant type, so every extraction is structured, auditable, and reviewable before anything reaches a legal artefact.
- Physical isolationEach firm's data lives in its own MongoDB database. One firm's records physically never sit alongside another's; isolation is at the infrastructure layer, not a permission check in app code.
- Region-pinned pipelineVertex AI and S3 are both pinned to Australian regions, so interview audio and inference never leave the jurisdiction.
Key integrations
Built-in safeguards
Launched as The Fact Pilot, live at thefactpilot.com.au.
Shipped in ~12 weeks from brief to production launch (Sep 2025).
Replaced a fragmented toolchain with one workspace covering the full investigator workflow, from intake to delivered report.
Firms onboard with their own isolated database, branding, and usage envelope.
- Per-firm workspaces with investigator, reviewer, and admin roles
- Audio upload + AWS Transcribe with speaker diarization
- Structured AI extraction per participant type (Claimant / Insured / Witness)
- AI-drafted witness statements and investigation reports with human-in-the-loop review
- PDF + DOCX artefact generation with firm branding
- TOTP MFA, audit trails, and tenant-level usage caps
Innovations
Each firm gets its own MongoDB database. One firm's data physically never sits alongside another's. Isolation is enforced at the infrastructure layer, so a bug in application code can't cross-contaminate firms.
Every AI call is constrained to a pre-agreed output shape per participant type (Claimant, Insured, Witness). The model can't improvise fields into a legal artefact. Outputs are structured, reviewable, and safe to export.
Interview audio lives in S3 ap-southeast-2. Inference runs on Vertex AI australia-southeast1. Residency is wired into every hop of the pipeline, from upload to rendered artefact.
- Investigators spend less time wrangling tools and more time on the facts of the case.
- Firm principals get a compliance posture they can defend: on-shore data, isolated databases, MFA-gated access.
- The AI layer accelerates draft work without introducing fabricated detail into court-adjacent artefacts.
- Live at thefactpilot.com.au
- Multi-tenant from v1, so new firms onboard without a fresh deploy
- Insurance and legal investigation firms that need on-shore data, isolated tenancies, and AI that drafts structured artefacts without fabricating facts.
- Factual investigators who want a single workspace for intake, transcription, extraction, and report delivery, not a toolchain stitched across six tabs.