|
I am a machine learning engineer in San Jose, CA, building LLM systems and AI agents, computer vision pipelines and production ML infrastructure. As an AI Trainer Fellow at Handshake AI I build evaluation environments for frontier models. I am co-first author of Interpreting Latent CoT Reasoning as Dynamical Systems (arXiv 2607.09698, poster at the ICML 2026 FoGen Workshop). I hold an M.S. in Artificial Intelligence from WPI (May 2026, GPA 3.97). My work spans mechanistic interpretability of latent chain of thought, TensorRT-optimized 3D perception, and multi-agent LLM pipelines that ship. Before WPI, I spent two years as a Lead DevOps Engineer at Persistent Systems (Top Talent 2024, Bravo Individual Award), owning CI/CD and AWS/Azure infrastructure across 20+ production projects with a 95% deploy-speed improvement. That production instinct is the thing I bring to research: pipelines that do not break, checkpoints that do not get lost, results that are reproducible 28 out of 28 times.
|
|
|
NVIDIA GTC 2026 Honorable Mention Recognized in the Google Cloud x NVIDIA developer community for inference-optimization engagement across three completed learning pathways. |
Algoverse Fellow AI Research Fellowship, Feb 2026 cohort Latent-CoT mechanistic interpretability |
Top Talent 2024 Persistent Systems 2 yrs Lead DevOps across 20+ projects |
|
Bravo Award Persistent Systems Sustained DevOps impact |
Top Intern 2022 20,000+ new joiners, Persistent India 30% reduction in app loading time |
B.E. Honors AI & ML, PES Modern College Pune 2017 to 2022 |
Bug fixes with regression tests in ML infrastructure I use, found by running the packages on Windows and CPU. Merged so far:
| Project | What the fix does | PR |
|---|---|---|
| NVIDIA NeMo Agent Toolkit | A module-level import made seven built-in tools silently fail to register on a plain install | #2474 |
| PyTorch torchtitan | download_hf_assets mutated its argument and raised an unreadable tuple error |
#5621 |
| Roboflow supervision | ComparisonAnnotator drew nothing on grayscale frames |
#2803 |
| A2A Python SDK | Malformed JSON-RPC error responses crashed the client instead of raising typed A2A errors | #1375 |
| NVIDIA Isaac Lab | configure_logging left the old log file open on Windows, so it could not be deleted |
#8809 |
| Hugging Face Hub | hf cache verify crashed when its output was piped on Windows |
#5255 |
Open PRs under review at Google (Flax, Optax, ADK), Microsoft (ONNX Runtime, ONNX Script, Olive, MarkItDown), Meta (ExecuTorch, TorchRL, torchao), NVIDIA (TensorRT-LLM), AWS (MCP servers), Anthropic (Claude Agent SDK), Hugging Face, MLflow, DeepSpeed, Ray and others: see the full list.
|
Reproduced CODI self-distillation and the COCONUT 4-stage curriculum as LoRA adapters on GPT-2, then ran a full DMD + PHATE + UMAP dynamical-systems analysis on 768-dim latent trajectories. Found opposite stability signatures. CODI behaves as a stable attractor and COCONUT as an unstable expanding system (DMD unstable-mode fraction 41% vs 85%), preserved across GSM8K and MATH.
|
Spatial probing of physical reasoning in Qwen2.5-VL, InternVL 2.5, and LLaVA-OneVision. Physics saliency maps across encoder, projection, and LLM layers, plus a 5-condition factorial LoRA ablation to find the leverage point for physical intuition.
|
|
Adapted FairEM (VLDB 2023) to annotator-level fairness. Discovered Non-Native annotators biased on hard questions (Fisher p=0.031, 6 of 10 FairEM metrics unfair). Robust across 5 cleaning strategies, 10K-shuffle permutation tests, BH-FDR-corrected.
|
3D perception pipeline for autonomous vehicles with TensorRT FP16 + layer fusion optimization, profiled via Nsight Systems and Nsight Compute. Composes MiDaS, YOLO3D, OSX, Mask R-CNN, and Detic into a Blender-rendered 3D scene.
|
|
CycleGAN + diffusion to synthesize paired healthy and diseased MRI and segment unsupervised. 65,072 BraTS2020 images processed with greater than 90% segmentation precision, without labels.
|
Confidence-aware LLM framework with dual-channel human-in-the-loop validation for ESI scoring: +25% triage accuracy, 40% fewer false escalations, 50% less miscommunication, 2.3x better patient-provider alignment.
|
|
Procedural game-asset generator combining Stable Diffusion + LoRA, CVAE-GAN, and Wave Function Collapse for spatially consistent biome maps. Unity prefab integration shipped.
|
Implements Mildenhall et al. NeRF (ECCV 2020) end to end. Trains on posed images, renders 360 degree orbits of Lego, Hotdog, and Ship scenes. Hierarchical sampling with positional encoding.
|
|
Auto-commits and pushes your changes with AI-generated commit messages via GitHub Copilot. Smart file filtering, debounced commits, status-bar control. Published on the VS Code Marketplace.
|
Multi-agent system using LangGraph + Amazon Bedrock RAG, AWS Step Functions, Lambda, S3, Textract. Production deployed: attorney review reduced from multi-day to ten minutes.
|
Languages and Core
Deep Learning, ML, Research
Inference, GPU, Performance
Cloud, DevOps, Backend
Web, Data, Tooling
|
ML Engineer Inference, Performance, Platform |
AI / Applied AI Engineer LLM systems in production |
Research Engineer LLM research, interpretability |
|
Forward-Deployed Engineer Research to customer outcomes |
Computer Vision Engineer 3D perception, multimodal, sensor fusion |
GenAI Engineer Diffusion, agents, RAG |
Locations: Bay Area San Diego LA Seattle Austin Dallas NYC Boston Remote-US, hybrid welcome in any major US metro
"My GPU runs hotter than my coffee." :: Dream. Hustle. Live.