Jonathan Tollefson
Design Engineer
Summary
Five years across marketing, UX, and enterprise data. I now build what I design: front-end, audio tools, and applied AI. Looking for a product team in audio, transcription, or AI.
Selected Builds
AI mastering agent · Python, pedalboard, Claude
2026 – Present
- Built an agent that masters a mix in a stateful, multi-step run, rendering candidates through FabFilter plugins.
- Shelved an active-learning model of my taste when blind tests showed it had learned the bias in my unblinded ratings, the known reward-model overoptimization failure.
- Every candidate is now loudness-matched (ITU-R BS.1770, true-peak limited) and judged blind against the untouched original, which won 26 of 58 votes (one rater, me, June to August 2026).
Audio catalog and search · Python, Whisper, librosa, Claude
2026 – Present
- A pipeline I wrote alone transcribes, fingerprints, deduplicates, and classifies about 20,000 audio files by content; a phone app finds any of them by a word.
Browser instrument · TypeScript, Web Audio API
2026 – Present
- Drum and bass patterns under keys played from the computer keyboard; it records takes as WAV files.
Experience
CHS Inc. · Business Analyst, AI & BI Engineering
Aug 2023 – Present
- Built agents on the internal AI platform, then designed and wrote the docs, guides, and onboarding chat that teach people to use them; other tools are adopting its design system.
- Built a Python tool that writes each dbt model's lineage to Excel, Power BI, and stakeholder briefs in one pass.
- Used AI to find data models missing semantic context and added it in Alation with their developers.
- Coordinate releases for the go-to-market data engineering team at a Fortune 500 agricultural cooperative; its AI and BI products serve thousands of employees, farmers, and growers.
Raylu, Inc. · Product Designer (Freelance)
Jan – Oct 2023
- Designed the brand, components, and prototypes for an AI startup; the founders carried them into meetings that closed a $4M seed, and engineers built from the same file.
Crediverso · UX Design Lead
Feb 2021 – Jan 2023
- Led the design system and onboarding for a bilingual banking app serving US Hispanic households.
- Ran usability tests with 50+ participants; the revised main flow cut bounce from 40% to 25%.
- Grew the pre-launch channel from 700 to 2,100 followers with A/B-tested messaging.
Education
Harvard University
2022
A.B. in Sociology, secondary field in History of Science
Dean's List (6×) · Rosenkrantz Discovery Grant
University of Minnesota
2016 – 2018
Transfer years; returned to Harvard to graduate
Skills
Build: TypeScript, React, Astro, HTML/CSS, Python, FastAPI, Claude Code · Design: Figma, Framer, Adobe Creative Suite
Models & Audio: Web Audio API, Claude, Whisper, librosa, demucs, pedalboard · Data: SQL, dbt, Power BI, Alation
Athletics Six-time state hurdles champion; set Minnesota's all-time 110m record and the University of Minnesota freshman record.
one run of the agent
- render up to 6 candidates, 4.2 on average
- blind test same loudness, labels hidden
- weigh votes become win rates
- build the master, from the winning takes
- decide
kept the untouched original 54%
42 of 78 runs that reached a decision
229 runs · 939 renders
log snapshot 2026-08-26 · one rater
the blind test's ledger · snapshot 2026-08-26
Each round plays two takes at the same loudness, unlabeled, judged on one axis; the untouched original is always one of them.
| date | axis | lufs | winner |
|---|---|---|---|
| 2026-06-10 | focal | -15.66 | render |
| focal | -19.56 | original | |
| dynamics | -21.37 | original | |
| space | -21.37 | original | |
| 2026-06-11 | space | -17.13 | original |
| focal | -16.53 | render | |
| space | -16.16 | original | |
| focal | -18.92 | original | |
| raw_vs_processed | -15.58 | render | |
| 2026-08-16 | focal | -16.30 | original |
the first nine of 50 rounds, and the latest
26 of 58
blind votes went to the untouched original
one rater (me) · 50 rounds in 6 sessions · 2026-06-10 to 2026-08-16 · no catch trials yet
the parser, on a sample dbt project
manifest.json parse 6 models · 6 edges · 3 layers
depth 0
depth 1
depth 2
pick a model to trace its lineage
writes to Excel Power BI briefs