Vantage AI
Production-grade lease analysis — web + Android — with PII redaction before any LLM call
The problem. Residential leases are dense, one-sided documents that most renters sign unread. Vantage AI — our UC Berkeley MIDS capstone — analyzes a lease PDF (web upload or Android photo capture) and returns a structured report: rights and risks, missing disclosures, cross-clause conflicts, and a negotiation plan, grounded in a California statute and guidance corpus.
How it works. PII is redacted (Microsoft Presidio + spaCy) before any text reaches an LLM. Eight lease aspects are analyzed concurrently through a retrieve → generate → judge loop over hybrid keyword + vector retrieval (Qdrant), with deterministic Risk and Disclosure engines layered on top so compliance findings never depend on model mood. FastAPI/asyncio backend, provider-agnostic LLM layer (LiteLLM), vanilla-JS web client, Kotlin/Compose Android app with on-device ML Kit OCR, deployed via Docker and Kubernetes.
My role. I owned the evaluation harness, the Android app, the compliance rules, and the legal corpus (33 merged PRs of the team’s 86). The eval harness is the part I’m proudest of: a three-arm benchmark (deterministic vs. RAG-LLM vs. no-RAG baseline) over a 58-case labeled dataset spanning 7 California lease templates with cross-template trap replants, scored with seeded bootstrap confidence intervals and exact McNemar tests.
Results — including the honest ones. From the 2026-07 three-arm run: zero hallucinated evidence in both pipeline arms; retrieval hit@1 of 0.60 vs. 0.52 for the no-RAG baseline (McNemar p = 0.0003, n = 218 — though baseline-dependent: p = 0.089 against an older model’s bare arm); trap recall 1.00 for the RAG arm vs. 0.25 deterministic; compliance precision 1.00; median latency 5.8 s deterministic vs. ~39 s LLM. We report the negative findings too — the LLM lease-review layer over-flags on clean templates (0/43 trap precision there), which is exactly the kind of thing an eval harness exists to catch.
Team. Built with Evan Powell, Terra Jiang, Trenton Carlson, and Adam Valadez.
Links. Official capstone listing (UC Berkeley I School) · CA Code Downloader — the statute-acquisition utility I built for the legal corpus. The application source is private.