Résumé
Seth Pratt
Applied AI Product Leader · San Francisco · sipratt@gmail.com
Summary
Applied AI product leader, currently on NVIDIA’s GenAI platform, with a decade of product work at NVIDIA, Lyft, Intuit, Meta and Visa. Outside work I build and measure AI systems: open benchmarks for clinical scribes and coding agents, large open models served on workstation GPUs, and a small studio of AI agents across five machines.
Experience
NVIDIA
2022 — presentApplied AI Product Leader
AI Research, enterprise AI, developer platforms, and autonomous vehicle validation.
Lyft
2019 — 2022Senior Product Manager
Level 5 Autonomous Vehicles · Universal search, developer productivity, and data infrastructure.
Intuit
2017 — 2019Product Manager
Search, small business finance, and experimentation platforms.
Meta
2015 — 2017Product Management
Formerly Facebook. Consumer growth, experimentation, and analytics tooling.
Visa
2014 — 2015Associate Product Manager
Payment gateways and enterprise product delivery.
Amgen
2013 — 2014Project Manager
Research operations and medical database resource planning.
Independent work
Open benchmark for ambient clinical documentation: 7 speech recognizers, LoRA, a claim verifier (98.7–99.6% injected-error recall) and an overnight search loop.
A four-tier measurement suite for AI-generated video; its first 59-clip sweep caught 26 extra-speech failures that file checks missed.
82 scored runs of local models and agent harnesses; found and fixed a grading bug that scored an answer 17/100; the same answer, re-scored, got 97/100.
Recipes for 100B+ open models on 2× RTX PRO 6000 and a Mac Studio: NVFP4/FP8, offload, MTP speculative decoding.
A local, encrypted AI assistant with private search, tested end to end with packet captures; 584 tests.
Everdust (Vision Pro / iPhone, App Store release pending), Create, ElmoSEO (live, paying customers), Cue, AutoTalent, Dr. Grey.
Tools
Evaluation: golden sets, LLM judges, claim-level verification, injected-error tests, paired bootstrap. Models and serving: vLLM, MLX, NVFP4/FP8, LoRA, speculative decoding. Speech and video: Whisper, diarization, SyncNet, LTX, H3. Agents: Claude Code fleets, MCP. Simulation: CUDA, WebGL/WebXR, Metal.
Education
University of California, Berkeley degree and year to confirm