TshepisoJafta/LexFlow

Voice-to-billing intelligence for South African legal professionals. FastAPI + React + Supabase + Gemini. # LexFlow: AI Billing Intelligence for Legal Professionals **Voice note in, reviewable billing entry out. Built for South African legal practice, with accuracy that is measured, not asserted.** **Live Demo** · **Published Evals** · **Architecture Docs** > **Recruiters:** click **"Try the demo"** on the login page, no signup required. The demo account seeds itself with sample data. --- ## What This Is LexFlow converts voice dictations into structured billing entries for legal professionals. Dictate after a consultation, on the road, or over WhatsApp: LexFlow extracts client, matter, duration and billable amount with Google Gemini, scores its own confidence, and presents an editable review form for human verification before anything reaches the ledger. AI output is treated as a draft, never as a fact. ## Measured Accuracy Extraction quality is benchmarked against hand-labeled golden sets of clauses from public documents across three families: a **South African scheme of arrangement under s114(1)(d) of the Companies Act 71 of 2008**, a US merger agreement, and a **payment-services agreement** carrying the fields contract-lifecycle workflows depend on (term end dates, renewal notice periods, liability caps). The benchmark's core metric is the hallucination rate: when a clause contains no value, does the model return null or invent one? | Metric (latest committed run, 33 items) | Value | |---|---| | Field-extraction accuracy | 100.0% (330/330 fields) | | Hallucination rate | 0.0% (0 of 316 null fields) | | Clause-type accuracy | 90.9% (3 documented borderline calls) | Methodology, golden sets and per-item results: `evals/`. Reproducible end to end from public SEC filings. The review process has already caught one answer-key error where the model was right and the label was wrong; the correction is documented, which is the point of publishing the labels. --- ## Architecture | Layer | What It Does | Technology | |---|---|---| | 1. **Input Capture** | Voice n …