dedna-genome-copilot-61049385056.us-west1.run.app/
Privacy-first genome copilot on Google Cloud Run: VCF/23andMe upload UX, interactive linear genome browser, grounded Gemini chat with AI→genome navigation chips, 5-tier variant screening, and research-grade disclaimers. Current build is a Vite/React SPA with in-memory variant state and simulated upload processing — async VCF pipeline and igv.js integration are documented in the product roadmap.
Personal genomics tools can be opaque: users upload complex files and receive static reports with limited interaction. A useful genome copilot needs to connect chat answers directly to genome navigation while maintaining privacy and research-grade disclaimers.
DeDNA provides a privacy-first genome copilot with upload UX, variant panels, a linear genome browser, grounded Gemini chat, and AI-to-genome navigation chips so explanations can point users to concrete loci and variants.
The product must avoid medical overclaiming, keep genomic state understandable, support future VCF and igv.js pipelines, and make chat grounding visible enough that users can inspect what the assistant is referencing.
The interface turns the assistant into a navigation layer over genomic data rather than a detached chatbot. Structured navigation contracts let AI responses move the user to relevant genome regions and variant records.
This shows Stefan's range across genomics UX, grounded AI chat, privacy-aware data handling, React/Vite product execution, Google Gemini integration, and responsible bioinformatics communication.
DeDNA treats personal genomics as an interactive evidence-navigation problem. The assistant is useful because it can point users to loci, variants, and findings instead of answering from detached natural language alone.
The key interface contract is navigation from AI output into genome state. Variant chips, browser focus, and finding drawers stay synchronized so explanations remain connected to concrete records.
The assistant must distinguish literature-grounded answers from user-uploaded variant analysis. That prevents the model from inventing variants or mixing public biomedical context with private genome state.
Personal genomics requires careful language around VUS, pathogenicity, rarity, hereditary risk, and non-medical use. The product demonstrates how to make uncertainty visible without making the experience useless.
The three-panel structure gives chat, genome browser, and findings distinct roles. This lets the assistant act as an interpretive layer over the genome rather than replacing the visual inspection workflow.
Typed variant state, navigation contract, and Gemini orchestration in a Vite/React SPA.
Three-panel workspace — chat, genome browser, findings — with synchronized locus navigation.
Dual routing: Flash + Search for literature vs Pro + thinking for uploaded-variant analysis.
Custom linear genome map with gene tracks, variant density, and coordinate navigation.
Grounded Q&A with structured sections: summary, findings, evidence, confidence, next steps.
Single API syncing chat variant chips, viewer focus, and detail drawer.
Search-grounded Flash (gemini-3-flash-preview) for literature; deep-reasoning Pro (gemini-3-pro-preview) for user data.
Filterable P/LP/VUS/LB/B findings with rare disease and hereditary cancer tracks.
Never invent variants not in uploaded data; explicit VUS uncertainty language; not a medical device.