AI Software Engineer — production LLM systems, multi-agent architectures & edge-native AI
MS Data Science @ Stony Brook University · ex-Genpact (2.5+ yrs) · ex-ARMA AI Labs, CA, US
I turn LLMs into reliable, cheap, production systems : RAG pipelines, agentic workflows, LoRA fine-tuning, and evaluation harnesses that make model behavior measurable. Recently focused on edge-native AI (Cloudflare Workers, Pyodide/WASM) that runs fast and costs near-zero.
| Project | What it does | Stack |
|---|---|---|
| ParamMem | Parametric reflective memory for LLM agents: cut agent reflection latency 94% (45s→2.8s), +11.4% HumanEval | LoRA · Llama-3 · Groq · ChromaDB · FastAPI · React |
| PrepGenie | Edge-native AI interview coach + resume reviewer; recursive LLM failover, LLM-as-Judge scoring | Pyodide/WASM · Gemini→Groq→Mistral · Gradio |
| AI Pentest Scanner | Autonomous multi-agent security scanner: recon → vuln detection → auto-report | Multi-agent · Groq/Gemini · Nmap/Shodan · React/Flask |
| Real-Time Stocks (Serverless) | Global stock dashboard on the edge; pull-on-demand caching → 1 API call/min/symbol at scale | Cloudflare Workers · D1 · R2 · Vite |
🤗 Also on Hugging Face — 15+ live Spaces including Neural Style Transfer (62 ❤️), plus PrepGenie, the AI Pentest API, and the ParamMem reasoning dashboard.
I contribute fixes to the backend and developer tooling I use day to day. A few of the most significant ones are in Supabase — auth now returns false instead of panicking on malformed SIWE signatures, and storage allows max-snippet -1 in x-robots-tag — Grafana k6, where I avoided a division by zero in counter rate thresholds, Cloudflare workers-sdk, which now rejects null observability config instead of crashing, Zendesk maxwell, where I fixed parsing of escaped backticks in quoted identifiers, Sourcegraph go-diff, where I fixed a panic on short diff --git args, Prometheus client_golang, where I fixed a panic on malformed remote-write label refs, Temporal sdk-java, where I fixed setRetryOptions merging initialInterval into congestionInitialInterval, Redpanda connect, where MySQL CDC now accepts unicode table names, and minitap-ai mobile-use, where I added LangGraph 1.2 / langgraph-prebuilt 1.1 support in ExecutorToolNode, plus ongoing work across agent-evaluation and MCP projects.
AI/LLM: PyTorch · TensorFlow · LangChain · RAG · LoRA · MCP · DeepEval · Agentic Workflows · LLM-as-Judge Backend: Python · Java/Spring Boot · C++ · FastAPI · Flask · REST · SQL Data & Infra: ChromaDB · Pinecone · Pandas · Spark · AWS · Cloudflare · Docker · Git
- Generative AI Engineer Intern — ARMA AI Labs (2025): built a DeepEval-based eval framework (7 metrics) + a production model-routing pipeline.
- Senior Associate Software Developer — Genpact (2022–2024): shipped Salesforce/Spring Boot systems for High Profile Clients; cut manual data entry 60%.
Open to AI/ML Software Engineer roles · Always happy to collaborate on LLM & agent projects.