Selected work
Agentic Graph Composer
Visual canvas and CLI for designing and editing LangGraph agent flows, with crash-safe run durability via a Postgres-backed checkpointer, OTel tracing for run visibility, and an eval harness that catches regressions before they ship. Published as a pip package on PyPI.
Dev Intelligence Platform
Multi-tenant, GitHub-based developer-intelligence platform surfacing PR bottlenecks, review health, and CI reliability from deterministic flow metrics — with classical ML for revert-risk and GenAI confined to grounded narration and clustering.
FieldOps Copilot
AI triage over NYC 311 service requests: a daily Socrata feed flows through embedding-based dedup (pgvector), a calibrated cascade classifier (cheap Groq tier → OpenAI) with a confidence gate, and a single LangGraph agent loop on only the ambiguous tail — which drafts human-gated work orders. An LLM-as-judge benchmarks the agent against a baseline with stated CIs.
Blitz — answers to your NFL questions
Agentic RAG over NFL play-by-play and roster data — a LangGraph router/retrieval/generation/reflection graph backed by ChromaDB, with a model-agnostic chat layer(Anthropic/OpenAI/Groq. Every LLM call is OTel-traced end-to-end.
Grounded RAG over Wikipedia articles
Grounded RAG over 1K Wikipedia articles — hybrid dense+sparse retrieval, cross-encoder reranking, cited generation, and an explicit abstain path when evidence is weak. Every answer is grounded, scored, ACL-aware, and designed to scale toward 10M docs by swapping infrastructure.
Experience
Principal Developer / Tech Lead — PinnacleCode · TOCA Football
Led a 6-engineer team delivering a sports-tech platform across 25+ studios — $4.3M+ in Stripe payments, −18% session failures, real-time booking over a shared Pub/Sub backplane.
Associate Manager / Sr. Developer — UNIRAC
Led UBuilder, a solar design & proposal platform for 500+ customers; cut design-to-proposal cycle by ~8 days and scaled the India team from 3 to 10.
Senior Software Developer — BrightWhite Innovations
Built a secure, auditable bidding repository with state-machine RFP lifecycle, vendor clarifications and immutable audit trails.
Software Developer — Amdocs
Architected an async ingestion pipeline syncing telecom bundling, promotions and regional tax logic into e-commerce checkout flows.
Software Engineer Intern — NestAway
Developed a lightweight Settlement Reconciliation service that used background workers to scan daily bank settlements against our internal financial ledgers, flagging payment mismatches or double-charges.
Capabilities
Writing
How I Published My First Python Package on PyPI
Publishing agentic-graph-composer taught me two things worth writing down: skip API tokens for Trusted Publishing, and if your CLI ships a frontend, bundle it into the wheel instead of making users install Node.
6 min
Sandboxing Your AI Agents: Docker's Cage for Claude Code, Codex, and Cursor
Docker Sandboxes run Claude Code, Codex, and Cursor inside real microVMs. Here's why that's what finally let me leave agents running unattended.
6 min
When Classical ML Beats LLM
LLMs are not always the right tool. A practical guide to the cases where a gradient boosted tree, a logistic regression, or a simple anomaly detector will outperform a large language model—and cost a fraction of the price.
11 min
How I'd Design a Multi-Tenant Booking Platform From Scratch
A general pattern for multi-tenant booking platforms: pre-materialized slots, pessimistic locking at checkout, Stripe webhooks as the source of truth, and an append-only credit ledger, distilled from a real production system.
15 min
How Temporal can simplify your microservices mesh
A practical look at moving orchestration out of a chatty service mesh and into durable workflows.
9 min
Redis for Everything: A Fun Exercise in Building an E-commerce System
A thought experiment in pushing Redis past caching — using it for carts, inventory locks, search, feeds, sessions, and rate limits to see how far its lesser-known data structures actually stretch.
12 min