Veritas AI

Inspiration

Legal work still starts the same way: scattered notes, screenshots, emails, call transcripts, and half-remembered timelines then a deadline. I wanted a system that could turn that chaos into a clean, auditable case record and keep momentum without sacrificing control or trust.

## What it does

Veritas AI is a lawyer agent built with ADK that organizes intake into a structured, versioned case file, then supports research, risk evaluation, drafting, negotiation language, and compliance checklists. It can also store case materials in a private corpus for later retrieval, and it requires explicit approval before any external actions like exporting to Google Docs or sending email.

## How we built it

We designed a case-first architecture: a persistent Case State Object with append-only versioning, schema-validated outputs for each stage, and workflow tools to track progress. Amazon Nova foundation models on Bedrock (via LiteLLM) power the reasoning and drafting, with an optional Nova multimodal path for extracting text from document images.

## Challenges we ran into

Getting “helpful” automation to stay safe was the hardest part—preventing accidental multi-step chaining, avoiding ungrounded citations, and enforcing approvals for anything that leaves the system. We also had to balance strict schemas (for reliability) with enough flexibility to handle messy real-world intake.

## Accomplishments that we're proud of

We built a system that treats legal work like an auditable process, not a chat: structured case memory, version history, grounded research support, and approval-gated actions. It’s practical enough to run locally or in a container and extensible enough to add more specialist capabilities.

## What we learned

A conversation isn’t a case file, state and versioning are non-negotiable. Trust comes from grounding and traceability, and real workflow value comes from doing the boring parts (structure, retrieval, checklists, approvals) consistently every time.

## What's next for Veritas AI

Stronger jurisdiction-aware retrieval and authority tagging, deeper multimodal intake (more document types, better extraction), richer integrations for case ops (while keeping approvals), and more robust evaluation so outputs stay consistently grounded and stage-to-stage reliable.