Co-founder & CTO, Remyx AI — experiment orchestration for AI teams.
Open source contributions across spatial-reasoning VLMs, agentic research automation, generative & multimodal AI, computer vision, robotics, and ML infrastructure.
The journey — from edge computer vision to agentic research automation (2019 → 2026)
🤗 283K+ downloads across 70+ open models & datasets (@remyxai — 21 models · 50 datasets) · ⭐ 1.4K+ GitHub stars across open-source projects
 ActionAI — real-time multi-person, spatio-temporally localized activity recognition from body keypoints, on the edge |
 nls-lumpy-torus — exploring interaction of geometry and dispersion to observe new blowup phenomena |
VQASynth — the open pipeline that turns raw images into spatial-reasoning datasets: CLIP retrieval → RAM/LLaVA captions → GroundingDINO/CLIPSeg → SAM → ZoeDepth → RANSAC planes.
Outrider — brief-to-PR research automation: arXiv → rank & license-enrich → select & preflight gates → draft → fidelity / convention / test audits → PR, with automated + human refinement loops and full run telemetry.
Fine-tuned VLMs for quantitative spatial reasoning (distances, sizes, directions — for robotics & embodied AI) and the synthetic-data pipelines that train them.
- SpaceQwen2.5-VL-3B — grounded spatial VQA · 150K+ downloads, the flagship
- SpaceThinker-Qwen2.5VL-3B — a reasoning spatial VLM (test-time compute) · 30K+ downloads
- SpaceOm · SpaceLLaVA (13B) · SpaceMantis · SpaceQwen3-VL-2B-Thinking
- Datasets: SpaceThinker · Robo2VLM-Reasoning · OpenSpaces · SpaceJudge — spatial VQA + reasoning traces
🗣️ The SpatialVLM research community credited remyxai for the open-source data-synthesis pipeline behind these models.
Often drafted with Outrider, then verified and refined:
- huggingface/peft — Riemannian-preconditioned LoRA optimizer (merged ✅)
- huggingface/peft — Super-Tuning: a new PEFT method (in review)
- DCDmllm/InstructSAM — inference fix + native C++/ggml runtime docs (merged ✅)
- huggingface/diffusers — HRDiT: training-free 4K high-resolution generation for FLUX.1-dev (community pipeline, in review)
- in flight toward LeRobot and TRL
- 🏆 Awards: #TFWorld TF 2.0 Challenge — Winner (Everybody Dance Faster — real-time motion-transfer booth, EdgeTPU + TF 2.0) · NVIDIA AI-at-the-Edge Challenge — 2nd prize (Saving Bandwidth with Anomaly Detection) · Snips.ai Smart Homes on the Edge — 2nd place (2019)
- 🎪 Hosted: Experiment 2025 — Remyx's inaugural event, hundreds in attendance
- 🎤 Conferences (2024–2025): ODSC West 2024 (LLMOps & MLOps track) · GitHub Universe · AI Agent Builders Summit — SF
- 🎙️ Podcasts: Generationship (Heavybit) — Ep. 20 Smells Like ML & Ep. 40 ExperimentOps · Adventures in ML (ML-149) · Code & Caffeine
- 📰 Press & features: NVIDIA “Meet the Maker” spotlight · The MagPi #80 (YogAI smart mirror) · Cerebral Valley — “ExperimentOps platform” · Cerebral Valley — “AI production assistant” (founder interviews) · NVIDIA “AI at the Edge” spotlight · SpatialVLM community shout-out
▶️ Demo: GitRank — Research to Testable PRs in Minutes
10+ years building production ML — Riot Games, Tubi (ranking & recommenders), Robust.AI (multi-camera 3D perception), and healthcare NLP (clinical IR & risk-stratification, presented at Stanford BMIR) — before co-founding Remyx AI. Along the way, a developer advocate partnering with NVIDIA, Intel, Arduino & the Raspberry Pi Foundation.
MA in Mathematics, UC Berkeley · BS in Mathematics, UNC Chapel Hill — where I studied trapping & blow-up phenomena for nonlinear PDEs on manifolds via finite elements, the same physics behind the nls-lumpy-torus gallery.
Remyx AI is an Alchemist Accelerator company (investors include Generationship & Roble Ventures).
remyx.ai · 🤗 remyxai · Docker Hub · smellslike.ml · Substack · Hackster · Sketchfab · YouTube · Discord · X · LinkedIn · Reddit