I build audio plugins in C++/JUCE, and the agentic tooling that develops and ships them.

Founder of Casa Occidente — three commercial plugins for macOS and Windows (AU + VST3), notarized, signed, and shipped. 13,000 users in the first year.

Most of my production code is closed source, so this profile is a route map rather than a mirror: the product is downloadable, the numbers are measured, and the engineering judgment lives in the repositories below.

A multi-agent harness I run in production against my own products — AI-assisted since 2024, my own harness since 2026.

• Orchestration. Sub-agents in parallel sessions, each with its own skills layer and its own scoped tool surface, working against a shared task tracker and linter.
• Permissions. Agents never hold my credentials. The agent proposes, the backend validates against a declared scope, and irreversible operations need explicit confirmation.
• Verification. Agents check their own output instead of trusting a clean exit — including rendering an artefact and inspecting the result.

Three public receipts. licensegate-bulk-cli, a TypeScript CLI that mutates real licences: 82 tests, CI on Node 20 and 22 (the run), dry-run first, a hard cap on how many licences one run may touch, and the API key never on the command line. The store's own agents.md gives agents a UCP/MCP path to browse and build a cart, and gates checkout on explicit buyer consent. The tool acts; the human authorises. juce-verify closes the other side: a harness that measures what a plugin claims you can hear — bit-exactness, zipper, block-size independence, deadline usage, latency — so a claim about audio becomes a number instead of an adjective.

dsp-probe — MIT, header-only C++17, zero dependencies. Extracted from RaiderPro, where it ran inside the DSP tests before being made standalone.

You cannot printf from the audio thread. At 48 kHz a 64-sample block lasts 1.333 ms, and nothing you would use to look inside a DSP fits in that window. So the numbers leave through an SPSC ring instead.

The ring is not a throughput buffer. It is a stall buffer: 8,192 rows is 10.9 s of audio at 750 rows/s, which is how long the draining thread can be dead before the probe starts losing rows. When it does, push() drops the incoming row and counts it. A probe that loses rows silently turns a passing test into a lie, so dropped() ships with the policy instead of as an afterthought.

18 test cases, 86 assertions, CI green.

• < 1.5 % CPU per instance, with negligible added latency, measured on the audio thread.
• Reported latency fixed at 5 ms — 240 samples at 48 kHz. Always the maximum: automating a lookahead would force the host to re-sync the timeline, so every user pays the maximum even at the smallest setting. Over-reporting is the safe direction.
• 20 % more clicks associated with a user, like-for-like, from a device-association algorithm I designed and shipped — measured against the existing attribution in BigQuery.
• 13,000 users in the first year, with support and the roadmap through my own hands.

Ten years of backend and platform before this, in Java, Kotlin and TypeScript:

• Mercado Libre, Technical Leader. A device data platform — collection → transformation → BigQuery over BigQueue (Meli's streaming platform), with per-use-case data products behind an authorizer — plus a Java library for device data generation shipped inside the Meli Android app.
• fabric.inc, via BairesDev. v2 of the order management system in Java: microservices, event-driven on SQS/SNS, MongoDB under production load.
• Lineage Logistics, via BairesDev. React pages and business logic, migrating a legacy FoxPro codebase serving inventory management across North America.

Generating options is cheap now. Everything below is about the part that isn't.

• A claim about what you hear gets a number. «This sounds better» is not a measurement, and an agent that says «it works» has told you nothing. The harness I use for a course on plugins reports a 14.7 dB click from an unsmoothed gain, 240 samples of latency, and 79.6 % of a block's deadline consumed by an allocation — and the first version of that last number was noise below the scheduler, which is written down too. A verification loop that never contradicts itself is not a verification loop; it is a rubber stamp.
• A missed buffer is a click, not a wrong value. The audio thread gets budgets and traces instead of assertions, and the target is fifteen instances in a real session, not a benchmark number on an idle machine.
• The real-time path and the control plane don't mix. Audio-path state changes lock-free; anything that touches the host or the UI goes through a thread hop, and never into sample-accurate DSP. Getting that boundary wrong is how plugins crackle under load.
• The tool should be able to leave the session. The Rider bakes its automation into the DAW, so you can delete the plugin and keep the mix. I hold the agent work to the same rule: the output has to survive without the thing that made it.
• A parameter range is a contract. Doubling the Rider's action range from ±12 dB to ±24 dB doubled the meaning of every existing automation lane in every host — a curve that reached +6 dB now reaches +12 — so the release notes say that in plain language, before anyone opens an old project and hears something different.
• When two answers are both right, hand the user the choice. RMS energy and LUFS short-term loudness measure the same take differently, so the plugin exposes both instead of picking one and calling it "the truth".
• An agent that can spend money must stop and ask. My store publishes its own agent instructions, and the payment step is gated on explicit buyer consent.

LinkedIn · diegortegaimi@gmail.com