Skip to Main Content

Codex

Built at Google I/O Hackathon Β· May 23, 2026 Β· San Francisco, CA

Codex β€” Demo video

🧬 Codex β€” the voice-conversational clinical variant interpreter Built for the Google AI Hackathon Β· Powered by Gemini Antigravity + Gemini Live Flash 🩺 The problem A patient gets a genetic test back. The report says: NM_007294.4:c.68_69delAG β€” Pathogenic That's it. Three cryptic tokens that determine whether someone needs prophylactic surgery, whether their sister should get tested, whether they qualify for a PARP inhibitor trial. Today, interpreting that result takes a board-certified genetic counselor 30–60 minutes of manual database hunting β€” across ClinVar, gnomAD, AlphaFold, PubMed, and ClinicalTrials.gov β€” followed by a hard human conversation. There are ~5,000 certified genetic counselors in the US. There are tens of millions of variants of clinical interest. The math doesn't work. πŸ“‰ ✨ The product Codex turns a raw HGVS string into a complete clinical briefing in under a minute β€” and then lets you talk to it. πŸ”¬ Drop in a variant β†’ an autonomous managed agent runs a full annotation pipeline in a sandboxed Linux environment: classification, population frequency, the AlphaFold structure with the variant residue glowing red, recent literature, open trials. πŸŽ™ Then ask anything in your voice. "What does pathogenic actually mean? Should my sister get tested? Can I still drink wine? What's a PARP inhibitor? Will my insurance use this against me?" β€” Gemini Live Flash answers in real-time, grounded in your analysis but able to draw on its full medical training for lifestyle, screening, and law questions like GINA Act protections. It's not just a search tool. It's the 30-minute genetic-counselor conversation, on demand, at 3 a.m. πŸ† Why this wins on the Gemini tracks πŸ€– Gemini Antigravity (managed agent) This is the perfect workload for a managed agent, and we lean into it hard: Real bioinformatics in a real sandbox. The agent runs Python in Antigravity's Linux VM, hitting MyVariant.info, NCBI E-utils (ClinVar + PubMed), AlphaFold-DB, and ClinicalTrials.gov v2 β€” live, every run. No mock data. Multi-step reasoning over heterogeneous sources. Resolve HGVS β†’ fetch ClinVar classification β†’ cross-reference gnomAD β†’ pull the AlphaFold structure β†’ render a lollipop plot with matplotlib β†’ score recent PubMed β†’ match active NCTs. The agent picks the order, handles 404s gracefully (e.g. BRCA2 exceeds AlphaFold's monomer limit β€” we degrade honestly), and bundles a manifest with inline-base64 artifacts. Cinematic replay for the demo. We capture step events to a step_events.jsonl log and replay them at 5Γ— speed on stage β€” so judges see the agent's reasoning unfold without waiting 90 seconds. The cache is our safety net; the live Run via Antigravity button is the credibility moment. 🎬 πŸŽ™ Gemini Live Flash (multi-turn voice) Live Flash is exactly the right tool for short, focused, scoped voice conversation β€” and we designed the UX around its strengths: Direct browser β†’ Gemini Live over WebSocket via the @google/genai SDK. No backend bottleneck. Native 16 kHz audio in, 24 kHz audio out. Persistent connection + mute toggle (Discord-style, not start/stop). One click to talk, one to mute. VAD handles turn-taking; we send audioStreamEnd: true on mute to flush cleanly. No teardown between turns means zero reconnection latency. System instruction is the magic. Each Live session is seeded with the variant's full analysis as system context β€” so it answers grounded, never hallucinates a trial ID. We deliberately split the prompt: variant-specific claims locked to the analysis, general medical knowledge unlocked (diet, NCCN screening, GINA Act, PARP inhibitor pharmacology). That's the difference between a useful counselor and a frustrating chatbot. Inline transcription for the deaf-accessible variant of the same conversation. πŸ’Ό Why this is a viable product 🎯 Real market Hereditary cancer testing alone is a $5B+ market growing 12%/year. Genetic counselor shortage is acknowledged by ASHG, NSGC, ACMG. πŸ₯ Real buyers Direct-to-consumer testing labs (Color, Invitae), hospital genetics departments, telehealth networks, pharma trial-matching platforms. βš–οΈ Compliance-aware "Decision support, not medical advice" is baked into every Live session. Citations are surfaced for every claim. Cascade-testing recommendations always route back to a human counselor. πŸ”Œ Composable The Antigravity pipeline is the moat. Swap the Live Flash front-end for a Twilio voice call, a Slack bot, or an EHR-embedded widget β€” the analysis layer is the same. πŸ§ͺ Scientifically honest Variants that don't fit AlphaFold's constraints (e.g. BRCA2 at 3418 aa) are reported as such, not faked. That builds clinician trust.

Team