.png&w=256&q=75)
🤓 Latest Submissions
LexPar AI
The Problem Independent lawyers handle most litigation but can't afford mock juries or senior-partner review. They walk into court never having heard their argument survive live, adversarial pushback — and have no way to prove trial-readiness to win premium clients. The Solution LexPar AI is a voice-driven courtroom flight simulator. Privately, an attorney spars against an AI Opposing Counsel that cross-examines and objects mid-sentence, then an AI Judge delivers a spoken ruling and scorecard. Publicly, it's a client-conversion tool: a solo attorney runs a live 10-minute session in front of a prospect to prove preparation and justify a premium retainer — a sales motion no rehearsal tool is built around. Architecture & AMD Advantage A React frontend streams audio over self-hosted LiveKit WebRTC to Python workers and a FastAPI backend. Opposing Counsel's reasoning model, Qwen2.5-72B, runs on a single AMD Instinct MI300X via vLLM/ROCm; its 192GB of HBM3 fits the 72B model plus case context on one GPU (~174GB resident, 100% GPU under load). A provider-agnostic router makes inference a config switch, not a code fork: one env var moves Opposing Counsel between cloud and the on-box MI300X, one-line rollback. Three systems keep the loop live and grounded: a 3-tier objection classifier fires instantly on clear infractions via text matching; replies are verified sentence-by-sentence against a structured ledger — gating fabricated citations before ElevenLabs TTS, cutting time-to-first-audio ~50%; and dual-corpus RAG grounds every argument in the pleadings and verbatim court rules, with a per-ruling citation audit trail. Commercial Potential LexPar targets the B2B legal market. Closed APIs bill per-token for hours of continuous voice; open-weight models on AMD run on flat compute — a high-margin SaaS model, with privileged case files that never leave owned infrastructure.
13 Jul 2026
