I design and ship AI systems where model reasoning, tools, memory, evaluation, permissions, and product infrastructure have to work as one reliable system.
My work spans architecture, implementation, research, and technical product leadership. I am most useful when an ambitious AI prototype needs to become an observable, testable, safety-conscious product across web, mobile, and cloud runtimes.
Portfolio · Nex Copilot · Plato Scientific · ORCID · GitHub
- AI systems architecture: define boundaries across model routing, tool execution, memory, retrieval, evaluation, policy, data, and user experience.
- Agent reliability and safety: build approval gates, typed tool contracts, postcondition checks, observability, failure recovery, and evaluation loops into the execution path.
- Technical strategy: translate uncertain product and research goals into staged architectures, explicit tradeoffs, measurable acceptance criteria, and release plans.
- Research-to-product delivery: turn papers, models, and experimental workflows into interfaces and infrastructure that people can actually use and review.
- Cross-functional technical leadership: connect engineering, research, product, design, and operations while staying close to implementation.
An AI-native agent platform spanning product workflows, developer tooling, benchmarks, observability, and web/mobile experiences.
- Architected agent-runtime and product surfaces across TypeScript, React, Cloudflare, Postgres, and mobile clients.
- Designed around governed tool execution, traceable runs, evaluation, and recoverable failure paths rather than prompt-only behavior.
- Built the platform as a connected system: runtime, marketplace, benchmark, documentation, and user-facing application share explicit contracts.
A chat-first autonomous agent for iOS and web with cloud, realtime voice, and on-device AI runtimes.
- Integrated model routing, durable memory, tool admission, human approval, browser workflows, and generative interfaces in one agent architecture.
- Built on Expo/React Native, Cloudflare Workers, Supabase/Postgres, and native Apple capabilities.
- Treats safety and release engineering as product features: fail-closed action gates, prompt contracts, worker-size checks, schema-drift checks, and runtime verification are part of the delivery path.
A multi-agent research system that turns experimental context into literature-grounded analysis and publication-style scientific artifacts.
- Combines retrieval, specialist agents, structured analysis, reviewer loops, and evidence-aware writing.
- Connects research workflow design with reproducible software and publication-oriented outputs.
An MCP server for structured genomic and biomedical tool use across ClinVar, genes, exons, introns, literature, and precision-medicine workflows.
- Converts heterogeneous scientific sources into typed, agent-usable queries.
- Focuses on evidence provenance and non-diagnostic framing for high-trust biological workflows.
A proof-driven engineering loop that coordinates Claude and Codex through planning, persistent execution, review, and adversarial quality gates.
- Uses explicit run state, isolated worktrees, verification evidence, and security/performance review before completion.
- Explores how coding agents can increase engineering throughput without removing human ownership or quality controls.
My research interests include agent evaluation, long-horizon reliability, memory and context systems, scientific agents, genomics, retrieval, and human-in-the-loop AI.
I have three peer-reviewed publications connected to ORCID 0000-0003-2268-053X:
- Chemical complementarity between immune receptors and cancer mutants is associated with increased survival rates — Translational Oncology, 2021.
- Small heat shock protein 22 kDa can modulate tau aggregation and liquid-liquid phase separation — Protein Science, 2021.
- A genomic approach to delineating scoliosis in arthrogryposis multiplex congenita — Genes, 2021.
Additional research-oriented systems include NexVar, DeDNA, and AgentSwarm.
I use agents as an engineering multiplier, not as a substitute for ownership. I define the architecture, constraints, acceptance criteria, and evidence required for a change. Agents help with exploration, implementation, testing, and independent review; production work still has to pass explicit technical and behavioral gates.
That workflow is reflected in the systems I build: typed interfaces, bounded permissions, durable run state, reproducible evaluations, reviewable diffs, and observed runtime behavior matter more than raw output volume.
- Languages: TypeScript, Python, Rust, Swift, SQL
- AI systems: tool-using agents, MCP, model routing, RAG, structured output, memory, multi-agent orchestration, evaluation, human approval
- Application: React, React Native, Expo, Next.js, Node.js, Bun, FastAPI
- Data and infrastructure: Postgres, Supabase, Redis, Cloudflare Workers, Docker, CI/CD, AWS, Railway, Vercel
- Research: PyTorch, scientific retrieval, genomics, ClinVar, variant analysis, literature-grounded workflows