Agentic AI EngineerI build the infrastructure
I build the infrastructure
AI products run on.
Building at the edge with Artificial Intelligence
I join the teams early, when the team is still figuring out the process, the architecture, and whether the whole thing would actually work. So I don't just write code. I make product and engineering calls that have to hold up later.
Most of my work sits between language models and real software: understanding where model reasoning breaks, building systems that fail safely when it does, and shipping things that still work when users hit them in messy ways.
I care a lot about the boring parts: latency, error surfaces, cost, retries, observability, and all the small details that never show up in a demo but decide whether the product actually survives in production.
3
AI products shipped to production
4
agents in a single pipeline
700K+
LLM calls monitored in production
Artifacts that I have built
Build Artifact #2
Health Voice
A clinical voice scribe that runs on one Mac. A nurse talks, it transcribes on-device, figures out who said what, pulls the medical terms, drafts a SOAP note, and files it to FHIR after a human signs off.
MLX Whisper (large-v3-turbo + base.en)silero VADspeechbrain ECAPA (speaker ID)
Jun 26, 2026Open artifact ->
Build Artifact #1Enterprise RAG
A full enterprise pipeline for the doc search system based on semantic RBAC RAG.
QdrantElasticsearchSupabase (Postgres + Storage)
Jun 25, 2026Open artifact ->
Projects I have crafted
PyTorchTRLDAPOPRMPEFT / LoRAQwen3-8BOllamaFastMCP
- Reimplementation of the ICLR 2026 AgentFlow paper as a local Qwen3-8B Planner, Executor, Verifier, Memory loop. Grammar-constrained JSON planning, Tavily search, and a sandboxed Python + SymPy executor.
- Swapped the paper's outcome-only GRPO for DAPO plus a learned Process Reward Model (Qwen3-0.6B regression head, trained on 531 DeepSeek-judged step labels) to get dense per-step credit. TRL ships no dynamic-sampling stage, so I wrote one.
- Full pipeline on a single A40: trajectory collection, step judging, PRM training, 300-step DAPO LoRA on Qwen3-8B (bf16), GGUF export, Ollama serving.
- Evaluation is leakage-free & quantization-matched: trained in bf16, scored on the served GGUF. GPQA-Diamond moved 40.0% to 45.0% (n=100), a directional cross-domain gain from a planner trained only on AIME math. AIME24 held flat (n=30).
OpenAI SDKAnthropic SDKLangGraphOpenTelemetry
- Enforces budgets on an agent before it acts. Decimal-precise caps on cost, tokens, wall time, and tool calls, checked pre-flight, so a runaway loop halts before the next expensive call rather than after it.
- Per-tool circuit breakers, and a verifier retry loop that feeds corrections back to the agent under the same shared budget.
- OpenTelemetry GenAI spans on every protected call. Failures return typed RunResult objects instead of raising, so callers can branch on the reason.
- Adapters for LangGraph and the OpenAI Agents SDK. Existing agents wrap without touching their code.
FAISSSentenceTransformersPyTorchSQLitePydantic
- Semantic cache for LLM agents. Embedding retrieval proposes candidates, then a learned pairwise classifier decides whether reuse is safe. "Approve this refund" never returns the cached answer for "deny this refund."
- Ships a pretrained classifier (v2) trained on 16,576 labeled pairs across 9 domains. At equal recall it holds 30 more precision points than a tuned cosine-similarity baseline.
- FAISS index, WAL-backed SQLite persistence, implicit bad-hit detection from downstream signals, gated retraining, CI across Python 3.11 through 3.14.
AsyncIOLiteLLMPydantic
- Multi-agent pipeline framework for Python 3.11+ with no required dependencies. Sequential, parallel, conditional, and retryable steps share one typed StepContext.
- Lifecycle events, flat execution traces, human review gates, and JSON checkpoint/resume for runs that outlive the process.
- Optional LiteLLM-backed Agent with structured Pydantic outputs. Shipped through v0.5.0 on PyPI
AsyncIOOpenAI SDKAnthropic SDKLangChainTyper
- Scores agents on pass rate over repeated runs rather than a single exact-match assertion. Agents are stochastic, so one green run is a sample of size one.
- Traces tool calls, step counts, and timing. Behavioral assertions collect and raise at the end: call ordering, argument schemas, latency bounds.
- Adapters for OpenAI, Anthropic, and LangChain. Typer CLI emits JSON reports that gate CI.
Where I've worked
- Built Browzer's Chrome MV3 recorder + CDP-native browser automation agent, achieving 95%+ precise AX/DOM element capture with cross-iframe support, obstruction checks, and real mouse/key/upload execution.
- Built a smart streaming ReAct loop across FastAPI + extension with SSE tool execution, multi-tab orchestration, safe parallelism, abort/continue, and audit logs.
- Cut automation LLM spend by roughly 67% using compact recording traces, context-window compression, prompt caching, and model-routing across GPT-5, Claude Sonnet & Haiku.
- Shipped a zero-LLM replay engine: recordings run as variable-driven tool-call templates, with a stateful AI fallback that resumes mid-run on failure.
- Shipped self-healing docs that auto-repair on UI drift — Haiku→Sonnet diff triage, LLM-free replay of intact steps, and a CDP agent that fixes only what changed.
- Shipped core features of an AI-powered real estate platform using Next.js, Nest.js, GraphQL, Redis, and GCP.
- Built the AI knowledge base service using FastAPI, LangChain, and vector retrieval pipelines, powering customer-facing search workflows.
- Developed document-ingestion pipelines using Google Cloud Vision, XLSX processing, and BullMQ workers, enabling automated extraction of customer data from spreadsheets and scanned records.
- Automated containerized CI/CD infrastructure via Docker, GitHub Actions, and Nginx for reverse proxy/load balancing.
- Built a LangChain + pgvector knowledge base powering AI-assisted document search and retrieval workflows, improving query accuracy by 15%.
- Developed scalable data-ingestion pipelines using bulk CSV processing and Celery workers, reducing processing time by 40%.
- Engineered a production PDF generation system transforming structured AI outputs and dynamic JSON reports into enterprise-grade documents.
- Automated deployment of AI services using Docker, GitHub Actions, and AWS EC2, establishing reliable CI/CD workflows for production environments.
- Received a personal offer from the CEO to join HeroUI (prev. NextUI) after making open-source contributions.
- Resolved 10+ bugs & delivered 7+ feature enhancements in core components including Calendar, Table and Pagination.
Achievements
- Top 1% TypeScript Engineer GloballyAlgora
- International Youth Math Challenge Gold HonourIYMC
- Amazon ML Summer School 2025Amazon