Executor AI

The AI that prevents executors from making expensive mistakes — by building a live intelligence graph of the estate and running a true agent that alerts before probate deadlines and liability triggers are missed.

Built for the Hackathon @ Berkeley 2026 (24-hour build).

The problem

When someone dies, the executor — usually a grieving family member, not a lawyer — is personally responsible for administering the estate: probate filings, asset inventory, creditor notices, debts paid in the right legal order, taxes, and distributions. Miss a deadline or pay out of order and the executor can be held personally liable. Families who can't afford a probate attorney do this alone, spending ~180 hours and making expensive mistakes nobody warned them about.

Executor AI is the expert in their corner: it reconstructs the estate from its documents and tells the executor the next action before it costs them. California probate only, and never a substitute for legal advice — for attorney-judgment questions it says so plainly.

What it does

Sign in, create an estate, and upload a will, deed, bank statement, or creditor notice. Claude parses each into a live estate-state graph. An estate-aware chat (text + voice) answers questions grounded in your documents. A real agent — the DeadlineAgent — proactively reasons over California probate law and tells you the next action before a missed deadline costs you, and a second ResearchAgent watches weekly for probate-law changes. Generated letters and emailed alert digests close the loop.

Architecture

Polyglot, two services + shared Redis. Python is the brain, TypeScript is the experience, Redis is the memory.

web/  (Next.js + TypeScript)  ── HTTP / SSE ──▶  agent/  (FastAPI + Python)
  auth · dashboard · chat · voice                 auth · documents · RAG chat
  Deepgram · Sentry                               DeadlineAgent · ResearchAgent
                                                  letters · email · Phoenix + evals
            └──────────────── Redis (KV state + vector search) ────────────────┘

Stack

• agent/ — Python · FastAPI · Anthropic (claude-sonnet-4-6 across parsing, the agents, chat, and letters) · OpenAI embeddings · Pydantic · bcrypt auth · Resend email · Phoenix tracing + LLM-as-judge evals
• web/ — Next.js 14 · TypeScript · Tailwind · Deepgram · Sentry · Zod
• shared — Redis: KV estate state + vector search for document retrieval, behind a store layer that supports Redis Cloud (cloud path in use), Upstash, or in-memory backends

Repo layout

• CLAUDE.md — working instructions for Claude / coding agents
• project_overview.md — full design, data shapes, flows, demo
• hackathon_tracks_and_prizes.md — tracks & sponsors
• docs/project_structure.md — implementation folders, ownership boundaries, and placeholder contracts
• docs/database.md — Redis KV/vector contract and database setup checklist
• docs/workstreams.md — per-member start points and stable dependency boundaries
• team/ — per-member role briefs (Members 1–4)
• agent/ — Python service · web/ — Next.js frontend

Getting started

# 1. Copy env files (won't overwrite if they already exist)
make env

# 2. Install all dependencies (uv for Python, npm for web)
make install

# 3. Start both services — agent on :8000, web on :3000
make dev

# 4. In a separate terminal: seed the demo estate
make seed

Fill in your API keys in agent/.env and web/.env.local after running make env. Minimum to start: ANTHROPIC_API_KEY in agent/.env. The store defaults to STORE_BACKEND=memory, so Redis Cloud is optional for local dev; voice (Deepgram), email (Resend), and observability (Phoenix/Sentry) degrade gracefully when their keys are unset — voice and email return previews instead of failing.

Phoenix tracing sends Anthropic, OpenAI embedding, and custom agent spans to PHOENIX_COLLECTOR_ENDPOINT (defaults to http://localhost:6006/v1/traces). Set PHOENIX_API_KEY when using Phoenix Cloud; local Phoenix does not require one.

Team