Agentic life copilot across web and mobile surfaces — chat, rich interactive cards, durable automations, memory, connectors, voice, and a governed Aura-default model route through Vercel AI Gateway.
Governed agentic copilot spanning web, portal, and mobile — search, analyze, automate, and execute across Web3/DeFi with read-only / approval / execution modes, MCP tool gateway, E2B sandbox runner, on-chain intent engine, HITL approvals, and machine-readable execution receipts.
End-to-end autonomous scientist at discovering.app — experimental data → literature-grounded idea, methodology, analysis, and LaTeX paper. Domain-routed retrieval, citation validation, evidence matrix, reviewer-panel revision loop, Modal/E2B executors, reproducibility manifests, LangFuse observability. Dashboard: Next.js 15 + FastAPI SSE. arXiv:2510.26887.
Customer-facing sales operations analyst for sales leaders — routes questions to fast or analysis agents, queries read-only Postgres evidence, delegates to data/coaching/artifact subagents, runs sandboxed Python analysis, and persists dashboards, charts, PDFs, CSVs, tables, and images.
Filesystem-first framework for durable AI agents — agent.ts configuration, instructions.md, typed tools, skills, channels, schedules, sandbox workspaces, and subagents in a pnpm/Turbo TypeScript monorepo.
Agent-native programming language and Rust toolchain with capability-secured effects, JSON-first diagnostics, stable Patch IR, LSP/package tooling, structural maps, and AI-optimized edit-check-repair loops.
FireCrawl-backed competitive-intelligence agent — maps competitor websites, scrapes pricing/features/about pages, extracts structured company context, then produces positioning, pricing, feature, weakness, and battlecard analysis.
Full-stack variant effect prediction — NexVar DNA LM (8.8T tokens, StripedHyena 2, 7B on Modal H100) with delta-likelihood SNV pathogenicity scoring, ClinVar benchmarking, and interactive hg38/hg19 genome workspace on Railway.
Privacy-first genome copilot on Google Cloud Run — VCF/23andMe upload UX, interactive genome viewer, grounded Gemini chat with AI→genome navigation, 5-tier variant screening. Current SPA uses in-memory variant state with simulated upload processing; async VCF pipeline planned.
CLI harness for clinical/genomic LLM benchmarking — PubMedQA baseline, LoRA fine-tuning via PEFT, GPU-aware dry-run fallback, toxicity/hallucination/privacy heuristics, PHI scanning, and CI across Python 3.9–3.11.
Python GPT-5.5 multi-agent orchestration — 35 scoped sub-agents (direct, council, research, security), context router + token budgets, hierarchical compression, council + security review, final synthesis with decision memory (~74% fewer input tokens).
MCP server exposing ClinVar interpretation and high-precision gene/exon/intron/PubMed queries for genomic agents.
Safety layer for operational automation — typed plans, policy gates, approvals, dry-run-first execution, postcondition checks, and machine-readable audit bundles.
Production monorepo (Next.js, Expo, FastAPI, MCP, AI SDK, Supabase) and a shadcn/AI-SDK component registry for agent chat, tool-call cards, streaming UI, and approval flows.
