Ted Jenkins

Content Systems & AI Infrastructure — 25 years building the pipelines behind audio, brand, and knowledge products, and still shipping production AI/audio software solo today

Burbank, CA — remote | reclusestudios@gmail.com | linkedin.com/in/ted-jenkins-21a7b57b

I build the infrastructure that makes content and audio findable, usable, and trustworthy — at massive catalog scale, across multi-brand portfolios, and now inside AI-driven products. One thing worth knowing up front: I don't write code by hand. I direct AI coding agents, applying 17 years of software QA instincts — test design, bug triage, systems thinking under real production constraints — to get production-grade results most non-engineers couldn't. Below are four things I shipped recently this way, on current-generation tooling (Next.js 16, React 19, JUCE 8, Gemini 2.5, local LLM inference), followed by the track record that got me here.

Recent Product WorkSolo-built, 2026 — Recluse Studios

StemScribe — Stem Separation & Chord Transcription for Musicians

Local-first desktop app — CPU-fast, CUDA optional

Challenge: Learning a song by ear means isolating instruments and figuring out the chords yourself — tab archives only cover popular songs, and only imperfectly. And practicing often means finding (or being stuck with) a backing track that isn't in your key or tempo.
Approach: v1 shipped as a hosted web app (Next.js/React + FastAPI, queued GPU pipeline on RunPod, Stripe subscriptions). When rented GPU availability became the bottleneck, rebuilt it as a local desktop app: a JUCE 8/C++ shell with a frozen Python companion running demucs (6-stem source separation) and a BTC transformer for beat-aligned chord recognition across a 170-chord vocabulary — entirely local inference. The CPU-only build generates 60 seconds of audio in about 47 seconds; a CUDA variant bundles the GPU runtime for those who have one. Stripe subscriptions remain, gating full-length playback, pitch-locked speed control, and PDF chord-chart export. When Stable Audio 3 shipped, added a feature-flagged Generate mode: describe a style, key, mode, tempo, and length, and a self-hosted Stable Audio 3 Small-Music model renders a practice track that flows through the identical stem-separation and chord-recognition pipeline as an upload — same-day chords and stems for a track that didn't exist an hour earlier. A v2 pass extended Generate into a "riff in, band out" workflow: audio-to-audio and inpaint modes take a recording of the user's own playing and build a full band around it, then the same separator mutes the played stem so they can practise their part live over the generated rest — plus an inline refine loop that regenerates from an edited prompt at a held seed.
Outcome: A local-first production product: no cloud dependency, CPU-only generation fast enough for real use (47s to generate 60s of audio), subscription billing retained, and production installers (CPU and CUDA) ready to ship — architected and built end to end by one person.

JUCE 8 / C++Python / FastAPI companiondemucs BTC transformerStable Audio 3StripeLocal CPU + CUDA installers

Corrigit — Click-Locked Reference Files From Phone Demos

Desktop app — custom beat-tracking algorithms

Challenge: A singer records a rough acoustic demo on their phone — no click, tempo drifting naturally. To give the drummer something to play to, a bandmate currently has to relearn and re-record the whole song just to lay a click under it. The use case comes straight from a working musician's forum thread; the manual workaround is standard practice in remote bands.
Approach: Built a JUCE desktop app: import the phone take (m4a/AAC included via Windows Media Foundation), paint regions on the waveform, and export click-locked files without anyone re-recording. Each region is either Follow (the click tracks the demo's natural tempo drift; audio passes through untouched — for rubato intros) or Lock (fixed BPM; audio time-stretched to match — for verses and choruses). Wrote the beat/tempo tracker in-house — spectral-flux onset envelope, autocorrelation with onset-supported octave resolution, Ellis-style dynamic-programming beat tracking — after the obvious library (BTrack) turned out to be GPLv3. SoundTouch handles time-stretch (LGPL, dynamically linked). Exports three files: click-only, a hard-panned click/program stereo mix that removes the DAW requirement for the drummer, and a byte-for-byte copy of the original. No GPU anywhere in the pipeline; reuses the Recluse license server and Stripe.
Outcome: Six and a half hours from conception to the first working render. v1 is code-complete, and the whole point holds: nobody re-records anything.

JUCE 8 / C++In-house beat trackerSoundTouch Windows Media FoundationRecluse license serverStripe

GigSmith — AI-Assisted Setlist Generator

Live, commercialized SaaS — $10/mo

Challenge: Setlist order isn't arbitrary — it shapes audience energy and pacing across a set — but building a good one by hand, every time, is tedious and easy to get wrong.
Approach: Built a three-stage pipeline: performers import their existing song catalog, an enrichment step fills metadata gaps from vetted sources, and a deterministic transition engine sequences the set using cross-track metrics — rather than leaving pacing decisions to an LLM's guesswork. An early, deliberate example of constraining AI output with hard logic instead of trusting a model to reason about ordering on its own.
Outcome: Fully built and commercialized, live today — real product, real subscribers, held back only by marketing budget, not by whether it works.
In Development: BackstageIQ — AI-assisted stage-plot conflict detection for live sound — is being folded into GigSmith for a one-stop live logistics solution for bands and techs: setlists, stage plots, and conflict checks in one place.

Next.js / ReactTypeScriptStripe subscriptionsDeterministic transition engine

Vocant — Natural-Language Audio Effects for Music Producers

VST3 plugin + local AI companion service

Challenge: Dialing in a studio-quality effects chain means hunting through hundreds of parameters across a dozen plugins. Producers wanted to just describe the sound they want.
Approach: Built a VST3 plugin (JUCE 8, custom WebView UI) exposing 283 automatable parameters across 32 effect modules — dynamics, EQ, drive, amp/cab modeling via Neural Amp Modeler captures, modulation, pitch, and creative effects. Paired it with a local companion service running a local LLM (Phi-3.1-mini via llama.cpp, auto-probing CUDA/Vulkan/CPU) that turns a prompt like "60s pop vocals" into a full, dialed-in effect chain in 1–3 seconds over OSC — with an optional cloud-LLM fallback for anyone who'd rather not run local inference. Then built the commercial layer to match: a Laravel/PHP license server with Stripe checkout, RSA-signed license keyfiles, seat management, and a 14-day trial flow.
Outcome: A real, licensable commercial plugin — not a demo — where audio never leaves the user's machine by default, and the entire path from "type a prompt" to "sound is dialed in" is under three seconds.

JUCE 8 / C++Local LLM (llama.cpp)OSC Laravel 11 / PHP 8.4StripeRSA-signed licensing

Also Shipped

Aurit (HarmonicFlow) — Adaptive Ear-Training

Accessible web app for musicians

A full ear-training app — intervals, chord quality, and chord-progression recognition with adaptive difficulty that targets your weak spots. Built with real accessibility engineering, not an afterthought: ARIA live regions, colorblind-safe indicators, a vestibular-safe animation toggle, and complete keyboard navigation. Persistent stats and shareable score links round it out.

Next.js 15 / React 19TypeScript 5Tone.jsVitest

Tactical Soundscape — Audio Engineering, Gamified

Browser-based educational game

Challenge: Frequency and EQ intuition are core to audio engineering, and notoriously hard to teach through reading or passive tutorials.
Approach: Built a tower-defense game where frequency-targeted towers defend a mix against enemies modeled on real mixing problems — mud, harshness, sibilance — so players build EQ intuition by fighting it in real time instead of studying a diagram.
Outcome: A complete, playable game that teaches real audio engineering concepts — proof that technical depth and a sense of humor aren't mutually exclusive.

Mutatur — Polyphonic Audio-to-MIDI

VST3 / Standalone plugin

Challenge: Learning a part by ear means transcribing it yourself — polyphonic audio-to-MIDI, so you can hear the notes as MIDI while a track plays and capture them into your DAW.
Approach: Built a polyphonic audio-to-MIDI VST3/Standalone plugin on the Spotify Basic Pitch model lineage (ONNX Runtime feature extractor + three RTNeural CNNs, Apache 2.0 credits retained). Emits a MIDI stream at ~1 s analysis latency while accumulating the same notes into a capture buffer you drag out as a .mid file. Later additions: a standalone metronome (click, time signature, volume, count-in) and MIDI export grid quantize; migrated onto the shared RecluseUI JUCE module.
Outcome: The fastest Recluse project from "works in my session" to "you can download it" — a shipped plugin in a growing collection, and the audio-to-MIDI side of the transcription stack.

JUCEONNX RuntimeRTNeural Spotify Basic Pitch lineageRecluse license server

MACE AI Drum Models — R&D Partner, Tensorpunk

GAN-based model training, UX/UI, and product development for Tensorpunk's MACE AI drum machine/sampler. Two custom percussion-generation models shipped commercially, including one trained on classic Roland TR-series sounds.

DIFFRACTION

A dependency-free (NumPy only) generative audio system: a 64-line prime-ratio delay lattice that turns a seed sound into evolving spectral interference. Research/art project exploring timbre as an emergent property, not a preset.

TinyReaper / TrimReaper2

Free, open-source one-shot extraction tool for hardware drum machines and synthesizers — originally released as TinyReaper, now under an active C++ rewrite.

VCV Rack — ACIDized Loop Player

A module bringing ACID's tempo-synced loop format into Eurorack-style modular synthesis — built purely out of curiosity about where that 25-year-old format still has room to go.

Career Highlights

157% Active User Growth at Zero Paid Spend

Tonic Audio — Senior Marketing Manager, Content Strategy & Growth

Challenge: A lean music-tech startup needed to grow its active user base with no paid acquisition budget, in a competitive category.
Approach: Built an SEO-driven content strategy and editorial calendar targeting composition, production, and remote-collaboration use cases; developed creator education workflows and cross-channel campaign orchestration (site, email, blog); stood up performance dashboards to keep the whole program accountable to funnel data, not vibes.
Outcome: 157%+ active user growth in 12 months, entirely organic — plus de facto product management responsibilities as the content strategy shaped the roadmap itself.

Brand Voice System Across Seven Brands

Fender Musical Instruments Corporation — Copywriter & Brand Historian

Challenge: Fender's portfolio (Fender, Fender Custom Shop, Squier, Gretsch, Jackson, Charvel, EVH) needed one coherent voice across product copy, a national dealer network, and major campaigns — without flattening what made each brand distinct.
Approach: Authored and maintained the internal brand copy style guide and trademark reference archive; built Fender Academy's dealer training content translating deep product knowledge into accessible material; wrote B2B and B2C copy including Fender Play UX content and campaign assets.
Outcome: A voice and terminology standard still enforced across marketing, training, and product documentation for seven distinct brands under one house.

450+ Compositions a Year, 90+ Loop Libraries

Sony Creative Software (formerly Sonic Foundry) — Sound Developer & Composer, 2004–2016

Challenge: The ACID product family needed a constant stream of fresh, genre-spanning demo content to show off what the software could actually do — and a growing catalog of loop libraries and algorithmic-music tools needed real compositional judgment behind them, not just content dumped into a folder.
Approach: Composed 450+ weekly demonstration tracks across a wide range of styles for the ACID family. Owned composition, editing, and metadata management for 90+ ACID musical and SFX loop libraries and 3 Cinescore (algorithmic music generation) libraries. Directed content selection, sound design, recording, and remixing across 210+ projects, including editing and remixing 200+ major-label artists from Miles Davis to Nine Inch Nails to David Bowie. Built and maintained a large attribute-based catalog system for production music licensing.
Outcome: A body of work that shaped how a generation of ACID users understood what the software could do — and the same compositional instincts that now drive Recluse Studios' production work and Tensorpunk's AI drum models.

Leading a 30-Person Crew, Then Moving Into QA

Sonic Foundry — Chief Video Encoder / Software QA Engineer, 1999–2002

Challenge: Digital distribution partners — Yahoo, Sony, Warner Bros., ABC, Microsoft — needed massive volumes of video ingested, corrected, and encoded reliably. Later, Sound Forge's batch-conversion tooling needed that same rigor turned inward, on the software itself.
Approach: As Chief Video Encoder, led a 30-person crew supervising large-scale automated video encoding for major distribution partners. Moved into Software QA for the Sound Forge Batch Converter: built manual and automated test scripts, chased bugs across media file I/O, GUI, and codec performance, and developed pre-processing techniques and .wmv/.asf codec templates to improve post-compression audio fidelity.
Outcome: Codec templates still in production use by Microsoft today — and a QA discipline (test design, bug triage, systems thinking under real production constraints) that, twenty years later, is exactly what makes directing AI coding agents work: knowing what to test and where things break, even without writing the code by hand.