Altamont — AI Legal Workspace

Problem

A boutique real estate law practice was drowning in document review. Every matter meant re-reading contracts, correspondence, and filings to answer the same kinds of questions — what does this clause say, when did this come up before, what's the relevant precedent in this file. None of it was technically hard to find, but it was slow, and slow work doesn't scale with caseload.

The firm also handled sensitive client documents that needed redaction before sharing — a manual, error-prone process when done by hand, and a real liability risk if a name or SSN slipped through.

Approach

Altamont is a workspace built around a matter: upload documents, and the system extracts text, splits it into overlapping sections, and embeds each section for retrieval. When an attorney asks a question, a hybrid search finds the most relevant excerpts — both semantic and keyword matches — and the model (Claude or GPT-4o, attorney's choice) answers strictly from those excerpts, with inline citations pointing back to the source document and section. If the file doesn't contain the answer, it says so instead of guessing.

Redaction runs as a two-stage pipeline: a regex pass catches SSNs, EINs, phone numbers, and account numbers, followed by an AI pass that finds addresses and dates of birth regex can't reliably catch. The flagged text is then permanently removed from the PDF at the content-stream level using MuPDF — not painted over, actually gone — with the original preserved in protected storage in case it's ever needed again.

The drafting tool turns a chat session into a starting document. An attorney picks a template — Motion, Letter, or Agreement — and the system pulls the relevant chat history and cited excerpts into a template-specific structure (IRAC for motions, formal letter conventions, full agreement sections with signature blocks). Every sentence in the draft traces back to a cited source; the model is explicitly told not to draw on outside legal knowledge.

Outcome

The workspace gives the firm a single place to chat with case documents, redact sensitive material with confidence, and generate first-draft motions and letters without starting from a blank page. Every AI query and response is logged with the attorney's ID, matter, and model used, so the firm has a full audit trail for every AI-assisted step in a case.

Technical highlights

• Cited document chat: hybrid semantic + keyword search over case documents, with inline citations and explicit "not enough information" fallback
• Verified redaction: regex + AI detection, permanent content-stream removal via MuPDF, with re-read verification that flagged text is actually gone
• Source-grounded drafting: Motion/Letter/Agreement templates generated from chat history and cited excerpts only — no hallucinated case law
• Data isolation: queries hard-scoped by organization and matter; documents encrypted in S3, database on SOC 2 compliant infrastructure
• Full audit trail: every AI query logged with user, matter, query text, and model used