AI/ML researcher & systems builder
I build language models and the systems that run them — from a from-scratch Mixture-of-Experts LLM and inference engines to production clinical RAG. Currently researching VLM agents and contributing to CNCF Kubeflow.
🏆 Hackathon champion · 💰 $500 Modal research grant · 🎤 Led community ML workshops on the Agno framework
Author on GeoAgent (Findings of EMNLP 2026) — evaluates VLM geolocalization through embodied navigation (the agent walks through Street View) instead of one static image.
- 6 models · 1,200 locations · 100 cities · best median error ~5 km
- Larger action budgets can hurt: unlimited actions ≈ 56× the cost of no navigation for little gain
- A developed-vs-developing region accuracy gap that image quality alone doesn't explain
Interactive explainer · Site repo
Contributor to Kubeflow (CNCF) — 11 PRs opened, 8 merged across two repos.
- CI + supply-chain security: replaced pip-audit with OSV-Scanner for CVE detection, weekly version-bump workflow, hardened the scanner workflow
- MCP protocol conformance tests + tool-schema snapshot test; in-flight agent-skills package (
skill://resources) - SDK: Spark Connect service-name fallback fix
PRs → mcp-server · sdk
|
🧠 KT_GPT 992.68M-param Sparse MoE LLM from scratch — 141.12M active/token (85.8% sparse), DeepSeek-style MLA compressing KV cache 8.8×, deployed on Modal.
|
⚙️ KLLM LLM inference engine from scratch — Go orchestration + CUDA kernels, ~20× throughput from continuous batching, int4 quant, MoE grouped-GEMM.
|
|
🩺 MediQuery Production clinical RAG decision-support — hybrid BM25 + Qdrant retrieval, cost-aware dual-model vLLM, NLI faithfulness checks tuned for MIMIC-III.
|
🕸 netintel AI criminal-network analysis (SIH 2026) — spaCy NER → evidence-backed knowledge graph → centrality, Louvain communities, anomaly flags.
|
| Group | Tools |
|---|---|
| Core ML | Python · PyTorch · Transformers |
| Techniques | MoE · MLA · SwiGLU · LoRA · GRPO · RAG |
| Backend | FastAPI · Agno · WebSockets · PostgreSQL |
| Systems | Go · CUDA · C · Rust |
| Infra | Modal · Docker · Kubernetes/Kubeflow · W&B · Qdrant · Redis · vLLM |
| Frontend | TypeScript · React |
43 public repositories · on GitHub since August 2024