Applied AI / ML Systems @ C3 AI · San Francisco Bay Area
Production AI systems — agentic workflows, ML systems, research engineering. Forward-deployed work for industrial customers today; ML team at Goldman Sachs before that. MSCS, NYU Courant.
dynadust3r-unofficial — Stereo4D (CVPR 2025) describes DynaDUSt3R but releases no training code or weights. Rebuilt the pipeline: motion head, losses, multi-GPU training, and a streaming data path over ~4 TB — naive MP4+NPZ decoding starved the GPUs, so clips became WebDataset shards with geometry computed on the fly. 29 h on 4×H100; 0.098 m EPE3D on ADT. Weights and datasets released.
Agentic development tooling @ C3 AI (proprietary) — An orchestrator agent decomposes natural-language requirements into specialized subagents that generate C3 Reliability application and model components, validate the output, and iteratively correct it when invalid.
DiffuserV2 — Velocity-parameterized diffusion planning with MPPI sampling and segmented replanning. Matches or beats Diffuser (ICML 2022) on Maze2D at half the diffusion steps; +21% on U-Maze.
JEPA world model — Action-conditioned latent dynamics in 89K parameters. Spatial conv latents, VICReg, and an inverse-dynamics auxiliary loss that recovers the action from the latent difference — resists collapse while forcing the latent to encode dynamics. 4.40 MSE on normal probes.
Multi-agent planning — Latent dynamics + MPPI planning on Overcooked-AI. Built the Q-value network and training stack; merged upstream #1 · #3 · #4. Planning and control, not LLM agents.
RL RoboSoccer — First place, AI RoboSoccer. A2C; the win came from observation design, not the algorithm.
- pico-llm#23 — RoPE and NoPE positional embeddings, model refactor, parallelized TinyStories tokenization. +423/−115, merged after 12 review comments.
- pldm_overcooked_ai — Q-value network and training pipeline (+1085), 1D value-shape fix, MPPI planner fix.
- pytorch-lightning#4459 — one-line documentation fix.
C3 AI — Data Scientist (June 2026–), Data Science Intern (2025). Forward-deployed work on C3 AI Reliability and agentic development tooling. As an intern, built a zero-retrain P&ID symbol recognition pipeline (YOLOv12n → MobileNetV3 triplet embeddings → class-adaptive k-NN) at 93.8% mean accuracy across seven reference splits.
NYU Courant — MS Computer Science, 3.909/4.0. Student researcher, Fouhey AI Lab.
Goldman Sachs — ML team, Enterprise Technology Operations, 2021–2024. Internal LLM resolution assistant over Llama 3 and Mixtral 8x7B with PEFT, automating ~90% of knowledge-based tickets in a ~100k/month workflow. Distributed log analysis with a custom pattern-matching DSL over Celery — 11x fewer production alerts. RoBERTa classification and entity extraction for financial email operations at ~92%, replacing an LSTM pipeline.