World models, VLA, and simulation
A robot learning loop from demonstrations through learned dynamics, planning, and held-out-world evaluation.
See the embodied-AI roadmap →
Featured work
Robotics first, evidence throughout
Shipped interfaces and measured experiments are labeled separately from active builds.
Working local demo · packaging now
Robot data
Measured
Robot policy
Tracklet turns robot episodes into training-ready supervision
A browser-first annotation workspace for persistent tracks, masks, grasp points, and manipulation phases. It uses gripper proprioception for deterministic phase boundaries and a compact browser SAM model only where pixels require it.
Measured
Robot policy
Diffusion policy from scratch
A 262M-parameter manipulation policy compared against the published reference under the same PushT rollout seeds.
Planning prototype AutonomyGPS-denied motion planning
Exact distance fields, sphere-traced collision checks, and paired RRT-family comparisons. State estimation and closed-loop flight remain future work.
Training run complete Open-model systemsAn 8B Rust coder, end to end
License-gated data, full FSDP2 fine-tuning, a retention regression that stopped the launch gate, FP8 serving, and GRPO.
Measured Audio intelligenceWho said what, seconds behind live
Whisper plus open diarization, with WER and DER written from scratch and evaluated on both clean speech and real meetings.
Live at gradientsmith.ai Model evaluationgradientsmith scores models on tests they cannot argue with
A verifier-first evaluation and post-training platform. Deterministic sandboxed verifiers instead of an LLM judge, an adversary that mines its own counterexamples, cost-aware routing across models, and RS-SFT plus GRPO with a reward-hacking monitor.
Systems view
The loop I like to own
Useful ML work is a connected chain; an accuracy number without data provenance or deployment constraints is incomplete.
01 · Data
Observe and label
Episodes, demonstrations, audio, filings, health records, licenses, and timestamps.
02 · Model
Learn and simulate
VLA policies, world models, retrieval, fine-tuning, self-play, and credible baselines.
03 · Evaluate
Test the decision
Matched seeds, held-out worlds, point-in-time splits, calibration, safety, and abstention.
04 · Operate
Deliver efficiently
Kernels, serving, cost, observability, and an interface that makes uncertainty legible.
Active portfolio roadmap
What I am building toward
These are transparent implementation dossiers, not claims of completed products.
Frontier open-model cost lab
Dated choices across Kimi K3, Qwen3.6/3.7, and GLM-5.2, plus fine-tuning, data, serving, and quality-per-dollar gates.
Read the playbook →My own AI coding harness
Codex and Claude behind one task protocol with isolated work, trace replay, test gates, routing, and cost accounting.
Read the harness design →GPU kernels to accelerator architecture
LeetGPU work organized by memory behavior and fused LLM blocks, then connected to a quantized inference accelerator.
Open the roadmap →Market intelligence with provenance
SEC filings, macro releases, prices, event graphs, regime shifts, and leakage-safe temporal evaluation.
Read the system design →Personal health intelligence
Meals, labs, vitamins, activity, family history, and genetics, designed as transparent wellness support, not diagnosis.
Read the safety-first design →
Open to AI / ML / robotics systems roles
Looking for an engineer who crosses the stack?
I studied electrical engineering at UT Austin and computer science at Stanford. I am most useful where models, systems, and product decisions meet: embodied AI, multimodal learning, reinforcement learning, model infrastructure, or accelerated computing.