nexvar-frontend-production.up.railway.app/
Full-stack variant analysis platform integrating the NexVar DNA language model — trained on 8.8 trillion tokens across all domains of life, StripedHyena 2 architecture, 7B/40B params, up to 1M bp context — with a Next.js research workspace and Modal H100 GPU inference. SNV pathogenicity via delta likelihood scoring, ClinVar benchmarking, and interactive hg38/hg19 genome browsing. Live at nexvar-frontend-production.up.railway.app.
Clinical and genomic variant interpretation requires long-context sequence reasoning, benchmark awareness, and careful UI translation. Most generic LLM interfaces cannot connect model scores, genome coordinates, known variant references, and researcher workflows in one place.
NexVar combines a Next.js research workspace with Python/Modal GPU inference for a DNA language model, delta-likelihood scoring, ClinVar benchmarking, known variant comparison, and hg38/hg19 browsing.
The platform has to bridge frontend interactivity with expensive GPU inference, handle genome coordinate context, keep variant explanations honest, and present pathogenicity signals without overstating clinical certainty.
The key move is treating foundation-model genomic inference as a full-stack product surface: model endpoint, benchmark harness, variant UI, gene search, and evidence comparison are designed together.
This highlights Stefan's capability in bio-AI product engineering, genomic model integration, variant-effect UX, Modal/H100 inference deployment, benchmark framing, and clinically cautious communication.
NexVar is a full-stack genomic model product. The depth is in translating DNA-language-model scores into a cautious research workflow that respects coordinates, reference builds, ClinVar context, GPU cost, and clinical uncertainty.
Variant analysis depends on local sequence context. The system scores reference and alternate windows rather than treating a variant as plain text, which keeps model inference connected to biological sequence structure.
ClinVar comparison and benchmark jobs give the platform a way to reason about thresholds, ROC behavior, confidence, and disagreement between model predictions and known clinical assertions.
Gene search, variant input, known-variant browsing, and AI-versus-clinical comparison sit in one product flow so researchers can inspect results without jumping between notebooks, APIs, and database pages.
The product has to communicate pathogenicity signals without making diagnostic claims. That requires confidence language, benchmark framing, and clear distinction between research support and clinical interpretation.
Research workspace for gene search, ClinVar browse, variant input, and AI vs clinical comparison UI.
Typed genome-api client layer for UCSC, NCBI, ClinVar, and Modal inference endpoints.
Serverless H100 GPU inference wrapping NexVar StripedHyena 2 scoring on Modal.
Interactive hg38/hg19 gene exploration with UCSC-backed reference sequences.
Manual SNV POST → delta likelihood score + pathogenicity tier + confidence.
Batch ClinVar scoring with side-by-side AI vs clinical assertion view.
analyze_single_variant fetches 8192 bp window, scores ref vs alt on H100.
StripedHyena 2 DNA LM trained on 8.8T tokens — score_sequences(ref) vs score_sequences(alt).
Offline Modal job: 500 SNVs → ROC/AUROC threshold calibration vs ClinVar classes.