I build the backend infrastructure that makes AI work at production scale — event-driven agent pipelines, RAG-based document intelligence, and GPU inference serving that holds up under real load.

7+ years of production backend engineering across Go, Python, and TypeScript — now focused on the agentic AI layer: multi-agent orchestration, retrieval pipelines, evaluation loops, and LLM integrations that stay correct under load, retries, and partial failure.

🤖 Nova — AI Advertising Automation Platform TypeScript · Node.js (Fastify) · Python (FastAPI) · PostgreSQL (Prisma) · Gemini · Google Cloud Pub/Sub · Cloud Tasks · Meta Ads API

Seven specialized services — page generation, content, image generation, QA gating, CMS publish, ad launch, budget allocation — coordinated by an event-driven TypeScript/Fastify orchestrator: Cloud Tasks command queues out, Pub/Sub events back, sustaining 20,000–30,000 ads/day. Message-ID deduplication gives exactly-once event handling over at-least-once delivery, and a monotonic status state machine means duplicate or out-of-order events can never corrupt campaign state. An automated QA gate blocks bad creatives before any ad spend; a monitoring service benchmarks live creatives and feeds results back into prompt selection — a continuously running production eval loop, not an offline benchmark.

🎨 Yodayo — Stable Diffusion Inference, Go Backend Go · RunPod · AWS SageMaker · GLM-4 · DeepSeek V3 · Claude

Go backend serving a platform with 2–4M monthly visits (peaked at 12M). Dynamic request batching, concurrent worker pooling, and GPU scheduling keep RunPod/SageMaker workers saturated. pprof-tuned hot path eliminated lock contention and allocation overhead. Integrated GLM-4, DeepSeek V3, and Claude for streaming Tavern chat across 105,000+ community models. Same architecture powers Moescape AI.

📄 Finance Document Intelligence — RAG Pipeline Python (asyncio) · RAG · pgvector (PostgreSQL) · Claude (Anthropic API) · Streaming

End-to-end document intelligence for finance teams: PDF/DOCX ingestion → chunking → vector embedding → hybrid retrieval → grounded Q&A with source citations. Replaced hours of manual cross-referencing with sub-second answers anchored to retrieved passages, not model weights.

mcpkit · Go Zero-dependency MCP server toolkit for Go — JSON-RPC 2.0, pure stdlib, concurrent request handling, HTTP gateway, and built-in observability. No third-party packages required.

mcpc · Go Zero-dependency MCP client for Go — library + CLI. Connects to any MCP server via stdio or HTTP, usable as a package or standalone tool. Pure stdlib.

langgraph-research-agent · Python Multi-agent LangGraph research pipeline — typed StateGraph, hybrid BM25 + semantic retrieval with RRF, cross-encoder reranking, FastAPI SSE streaming, and LLM-powered synthesis. Full RAG stack with pgvector and Chroma.

finsight-crew · Python Multi-agent CrewAI crew that researches a stock and writes a cited investment brief — company analysis, market data, SEC filings, and risk assessment in one run.

extractflow · Python Schema-validated structured extraction from PDF, DOCX, CSV, and email using LLMs — built for automation pipelines.

flowra · Go Self-hostable workflow automation engine in Go. Define LLM-powered pipelines as a single YAML file, ship as one static binary.

LangGraph · CrewAI · MCP (Model Context Protocol) · LangSmith · pgvector