Hammurabi
An AI legal-analysis SaaS for Israeli lawyers, in Hebrew and English. I took it from first commit to production in about 3 months, then grew it into a paid product.
- Client
- Hammurabi
- My role
- Founding & lead engineer
- Team
- Solo for 10 months, then a small team
- Period
- 2024 – 2026
Results
~3 months
From first commit to productionThen paid tiers
10 months
Built solo before the team joinedThen led
~12
Legal AI tools on one frameworkShipped
4
Retrieval strategies compared in TruLensBest shipped
Context
Hammurabi is an AI legal-analysis SaaS for Israeli lawyers, in Hebrew and English. Lawyers upload pleadings, judgments and case files, chat with them, and run AI tools that analyse a pleading against a profile of the assigned judge. It sells on subscription tiers with credits.
The client, an Israeli founder, wrote the requirements. I turned them into architecture and code.
My role
Founding and lead engineer for the whole engagement, April 2024 to early 2026. I built it solo for the first 10 months, then led a small team through a PR-based workflow, and stayed the #1 contributor across ~2 years.
The problem
- Retrieval needs context. A legal answer depends on the text around the matching passage, in Hebrew and in English.
- Hebrew everywhere. Right-to-left UI, Hebrew prompts, and OCR for scanned documents.
- A growing tool list. Each new AI tool needed pricing, plan gating, background execution and a downloadable report.
What I built
Documents go through OCR and chunking, then Voyage embeddings into Pinecone. Retrieval feeds the LLM layer, and results come back as chat answers or tool reports.
- 01Ingest
- Pleadings, judgments, case files
- OCR for scanned Hebrew
- 02Index
- Sentence-window and linked chunks
- Voyage embeddings into Pinecone
- 03Retrieve
- Query classifier
- Vector search, Voyage rerank
- Context rebuilt around each hit
- 04LLM layer
- Claude, GPT, Gemini behind one factory
- Python FastAPI analysis services
- 05Deliver
- Chat answers
- ~12 tools with HTML, DOCX, PDF reports
- Credits and plan gating
- Platform. Next.js, tRPC, MongoDB and Redis, with single-seat sessions enforced in Redis, an admin console, and a case-management layer.
- Custom RAG retriever. I first ran a TruLens evaluation harness comparing naive RAG, sentence-window and auto-merging retrieval on Hebrew and English legal text. Then I designed a sentence-window + linked-chunk retriever: each chunk stores its neighbours’ ids, so after vector search and Voyage reranking the context is rebuilt around each hit. A query classifier routes simple, file-specific and contextual questions differently.
- AI tools framework. Each tool declares its credit price, minimum plan, background job and report format (HTML, DOCX or PDF). I shipped ~12 legal tools on it, including argument analysis, contradiction detection and success-probability scoring.
- Judge profiling. Profiles built from a judge’s past rulings, with scrapers that fetch missing rulings. Python FastAPI services, built with a teammate, run the heavier analysis with Pydantic structured outputs.
- LLM layer and billing. One factory over Claude, GPT and Gemini, with admin model switching and editable prompts. Stripe subscriptions and credits.
- Word add-in and v2. I led a Microsoft Word add-in that brings suggestions into the document as accept-or-reject changes, and architected the v2 rebuild on Convex, WorkOS and Mastra.
Results
- In production about 3 months after the first commit, then improved continuously as a paid product; the first 10 months built solo.
- ~12 legal AI tools on one framework: a new tool became a config entry plus a handler.
- A retriever chosen by evaluation against three baselines, not by intuition.
- Runtime Hebrew/English switching across 1,195+ translated call sites without rewriting them.
What I’d do differently
- Evaluate retrieval from the first version. RAG v1 shipped in spring 2024; the TruLens harness came that November. The harness led to the better retriever, so I’d build it alongside v1.
- Settle billing early. Billing moved from PayPal to WooCommerce to Stripe, and each move was real work. I’d push for the long-term provider before the first paying user.
- Design for organisations from the start. v1 was built around individual accounts. Multi-tenant organisations arrived with the v2 rebuild, when they could have been a v1 data-model decision.