I build backend systems that scale reliably and LLM-powered features that ground answers in real data. At OurFreedom.ai I led development of Carla, a LangGraph agent serving 1,000 requests per day, and designed the evaluation pipeline that cut hallucinations from 16% to under 5%. Before that I shipped RAG pipelines, connection-pooling fixes that reduced peak connections by 85%, and Kafka consumers that reconciled more than 40 million part records.
| Project | What it does | Impact | Stack |
|---|---|---|---|
| Carla agent | LangGraph tool-calling agent with Pydantic validation, retry logic, and human review queue | Serves 1,000 requests/day; hallucinations dropped from 16% to under 5%, tool-selection accuracy at 91% | LangGraph, Pydantic, Langfuse, AWS, Prometheus, Grafana |
| University Advisor Chatbot | Agentic chatbot delivering source-cited university recommendations through five custom tools | Integrated into study-abroad consultancy, serving 500+ students weekly | Python, FastAPI, Next.js, AWS Bedrock, Claude Sonnet |
| Dispatch Engine | Delivery dispatch system matching orders to drivers via Redis GEO and PostGIS, with Kafka state publishing | Handles 300 orders/day, reduced delivery inquiries by 90% | Java, Spring Boot, PostGIS, Redis, Kafka |
| Formulary ingestion pipeline | PDF-to-model pipeline with LoRA fine-tuning of Llama 3.1 8B, served on-device | Zero per-token cost, fully local inference | Python, MLX, LoRA, Docling, PostgreSQL, Llama 3.1 8B |
| RAG pipeline | LangChain RAG with pgvector grounding answers in 40,000 indexed documents | Hallucination rate from 18% to 4%, measured on 500-query weekly eval set | Python, Pydantic, LangChain, pgvector |
| Kafka reconciliation consumers | Idempotent consumers detecting revision mismatches across PLM and ERP systems | Resolved 1,200+ monthly mismatches across 40M+ part records | Java, Spring Boot, Kafka |
Carla agent, in depth
The approach: I built Carla as a LangGraph agent with Pydantic-validated outputs, retry and fallback handling, and a human review queue for flagged outputs. I designed an agent evaluation pipeline and a 150-case gold set run nightly with Langfuse tracing, then added Prometheus and Grafana monitoring for p95 latency, tool failure rate, and cost per request.
The result: Carla serves 1,000 requests per day. Hallucinations dropped from 16% to under 5% and tool-selection accuracy reached 91%.
University Advisor Chatbot, in depth
I developed an agentic chatbot orchestrating Claude Sonnet through a tool-calling loop across five custom tools on a FastAPI backend with a Next.js chat UI.
The chatbot serves more than 500 students weekly.
| Period | Role | Company | Highlights |
|---|---|---|---|
| Jun 2026 – Present | Software Engineer, Founding Team | OurFreedom.ai | Built Carla agent serving 1,000 requests/day; reduced hallucinations from 16% to under 5% and raised tool-selection accuracy to 91% |
| Jun 2025 – Aug 2025 | AI Software Engineer, Summer Intern | Tiny Archives | Shipped RAG pipeline grounding answers in 40,000 documents; cut hallucination rate from 18% to 4% on 500-query weekly eval set |
| Jul 2023 – Aug 2024 | Founding Engineer | Duende Ingenious | Introduced PgBouncer pooling, reducing peak connections by 85%; cut silent order failures from 0.5% to 0.3% and release cycle from two weeks to one |
| Aug 2021 – Jun 2023 | Software Engineer I | eQ Technologic | Built Kafka consumers reconciling 40M+ part records, resolving 1,200+ monthly mismatches; reduced failed sync jobs by 40% and eliminated 200+ manual manifests |
- Master of Science, Computer Engineering – New York University, May 2026
- Bachelor of Engineering, Electronics & Telecommunications Engineering – Vishwakarma Institute of Information Technology, May 2021
- locksmith — A migration linter that parses PostgreSQL migration SQL and predicts how long each statement blocks writes, using the lock matrix for the target server version…
- gridboard — A real-time dispatch console for a grid operator that downsamples under backpressure and marks panels STALE instead of showing old data as live.
- consentgraph — Identity resolution where per-purpose consent gates the graph join itself, with match quality measured as precision against planted ground truth and a…
- traceboard — Requirements traceability tool that links requirements, tests, code, and defects, and flags stale coverage where a test passes against a requirement that has…
- sysprobe — An eBPF profiler for Linux that records per-process syscall, run-queue and TCP connect latency, exported as Prometheus histograms and a top-style view.
- originbound — A sealed lab where one relay attack succeeds against TOTP and fails against WebAuthn, with the clientDataJSON origin field shown as the cause.
- alwaysvalid — A sequential A/B test calculator using always-valid inference (mSPRT, confidence sequences), shipped with an A/A simulation showing how daily peeking inflates…
- schemadrift — A GraphQL breaking-change gate that downgrades severity using real field-usage telemetry, naming the client that still calls a removed field, with a GitHub…
- adrlint — A decision-record linter that builds a graph of architecture decisions and flags unresolved contradictions, expired reviews and orphaned supersedes chains.
- dmarcwatch — Email deliverability monitoring that parses DMARC aggregate XML reports to separate mail that never arrived from mail that arrived and was ignored, and flags…