"I don't just wrap APIs. I build the evaluation harnesses, the CUDA kernels, and the 7,000-test CI pipelines that make AI reliable in production."
I am an AI & Research Engineer bridging the gap between cutting-edge LLM research and hardened production infrastructure. With 1.5+ years of remote experience on US-based teams, I specialize in Agentic Systems, RLVR (Rule-Based Reinforcement Learning), and GPU-optimized Inference.
Whether it's ranking 13th in Berkeley's AgentBeats Security Arena, cutting TTS inference latency by 4.5x via CUDA Graphs, or authoring 7,275+ Pytest functions to guarantee reliability, I ship end-to-end systems that scale.
| 🧠 AI & Research | ⚙️ Systems & Backend | 🛡️ Quality & MLOps |
|---|---|---|
| LLM Agents & MCP Tooling | FastAPI, WebSockets, AsyncIO | Pytest (7,200+ tests), CI/CD |
| GRPO / RLVR (TRL), LoRA | Docker, RunPod Serverless, AWS | OpenTelemetry, SigNoz Tracing |
| RAG, Agent Memory, Red-Teaming | CUDA Graphs, bfloat16, KV-Cache | Ruff, Mypy, 93% Coverage Gates |
| PyTorch, Transformers, Whisper | PostgreSQL (asyncpg), Redis, S3 | Execution-based Eval Harnesses |
|
Open-Source PyPI Package + MCP Server Makes Jupyter notebooks LLM-efficient. Agent-optimized output cuts tokens up to 80% across 8+ formats, with tiktoken budgeting and streaming for 10MB+ notebooks. 🏆 10,000+ PyPI Downloads View Repository → |
Adversarial Robustness Platform Scenario-agnostic attacker/defender agents. 8-layer defense pipeline covering OWASP ASI01-10. 2,295 tests at a 93% coverage gate. 🏆 Ranked 13th Overall (49.7% Win Rate) , 18th in Attack (14.4% Win Rate) and 6th in Defense (85.1% Win Rate) in Berkeley's AgentBeats (Lambda Agent Security) competition. |
|
Systems-Level TTS Engine Up to 4.5× faster inference (RTF 0.59 → 0.13) via bucketed CUDA Graphs, bfloat16, and KV-cache tuning. Zero-hallucination pipeline deployed to RunPod. |
Verified RL Training Stack Built the complete training + evaluation stack for a math LLM (GRPO on TRL). 22,796 deduplicated samples, oracle-verified gold solutions, sandbox pool sustaining 389 exec/s. 💰 Trained for < $100 |
Junior Data Scientist @ Teamlift (US Remote) Jan 2024 – Dec 2024
- Scraped and structured thousands of AI tools to power the Teamo recommendation engine.
- Built automated verification pipelines (HTTP checks, freshness validation) eliminating stale entries at scale.
- Worked directly with US stakeholders, shipping dataset improvements on tight iteration cycles.
Data Science Intern @ Teamlift (US Remote) Jul 2023 – Dec 2023
- Engineered robust scraping pipelines for dice.com with deduplication, rate limits, and retries.
- Trained time-series models on skill-demand signals to forecast market trends and inform product decisions.
I'm open to consulting and full-time roles in AI Infrastructure, Agent Security, and Applied LLM Research.
