John Hartquist

San Francisco, CA

• •Built Ruby and Node.js OpenTelemetry auto-instrumentation SDKs powering Foam's AI agents: fail-safe APIs, coexistence with APM vendors (Datadog, Sentry, New Relic), PII-redacted network and database capture, and 100% branch coverage.
• •Built cross-language (TypeScript, Python, Ruby) end-to-end verification harness, translating requirements into 1,700+ audited tests across relevant library and runtime versions; surfaced upstream bugs and patched cross-vendor conflicts.

Remote

• •Shipped features across TakeOne's ML video pipeline, processing 10,000+ phone-recorded takes from 1,500+ users: video matting and segmentation (SAM, Track-Anything, GroundingDINO), skin-tone correction, audio cleanup (DeepFilterNet + FFmpeg loudness normalization), and transcription (Whisper, Deepgram).
• •Built LLM video-editing pipeline with validation gates, LLM-as-judge, and retry-with-feedback.
• •Built the full-stack customer portal solo (TypeScript, Next.js), where client admins review and publish videos.
• •Built internal tools in Rust for bulk user onboarding, provisioning, data validation, and API testing.
• •Wound the company down in March 2026.

Kirkland, WA

• •Prototyped AR measurement and virtual-object simulation in Unity for in-store and at-home use cases.
• •Built React / Babylon.js inspector and ingestion tool for 3D home assets; published a custom open-source product viewer.
• •Built a room-scanning prototype using Apple's RoomPlan API to capture 3D room geometry.

Seattle, WA

• •Productionized a financial-reconciliation ML model (Python, XGBoost) and migrated the pipeline to Spark (Scala), achieving a 100x+ speedup over the prior Python implementation.
• •Built dataset, training, and evaluation tooling to iterate on reconciliation model accuracy.
• •Built a full-stack platform for running Jupyter notebooks in production (TypeScript, React, Next.js, Docker, GCP).

San Francisco, CA

• •Studied classic ML, deep learning, and reinforcement learning full-time for a year.
• •Built `fastai_audio` (194★), a library for on-the-fly GPU spectrogram generation during training; write-up.
• •Placed top 8% (39th of 556 teams) in Kaggle's Freesound audio-tagging challenge with CNN ensembles.

San Francisco, CA

• •Built licensing and brand-protection platform (Ruby on Rails, Angular) used by HDMI, Dolby, and Philips.

Mountain View, CA

• •Built iOS app (Objective-C) and Python/Django backend services for an early-stage 'Tinder for Jobs' startup.

San Luis Obispo, CA

• •Built internal services for the CreateSpace Royalties & Financials team (Java, Spring, AWS) using test-driven development.

• •Composes System 1 judgments into System 2 decisions in plain Rust, batching many questions per request.
• •Answers come back as the caller's own types, from derived enums to arbitrary structs, with calibrated probabilities intact.
• •Procedural macros validate literal questions at compile time; questions from data and every answer are checked at runtime.

• •Rust implementation of Alexandre François's Resonate algorithm: per-bin spectral analysis without FFTs or windowing.
• •Cross-platform: native on macOS, Linux, iOS, plus browser (WASM). Python (PyO3) and JavaScript bindings.
• •Outperforms the Apple-only C++ reference implementation by 1.6x on M2 Max.
• •Live WebGL streaming-spectrogram demo and write-up at johnhartquist.com/resonators.

• •Claude Code + Remotion starter kit for programmatic video; walkthrough video has 43,000+ YouTube views.

• •Converts rustdoc JSON into markdown so AI coding agents can query crate docs without spending tokens parsing HTML.

• •Runs tasks autonomously in isolated git worktrees with verification (tests, clippy, build) gating merges.
• •Retry-with-learnings: failed attempts produce structured findings that feed forward into the next attempt's context.

• •End-to-end Rust ports of CREPE, Swift-F0, and Spotify's Basic Pitch via Hugging Face's Candle framework.
• •Numerical-parity tested against the original TensorFlow / PyTorch implementations.

• •Cal Poly MS thesis (2012): real-time polyphonic guitar transcription using Non-Negative Matrix Factorization.
• •Full Rust rewrite, GuitarSet evaluation, neural experiments with HCQT features and Mamba-family architectures.
• •Adapted Spotify's Basic Pitch to run in real-time in the browser via `resonators` front-end.