Notes on shipping production AI | Evan Parra

Blog

Notes on shipping production AI.

How I think about replacing per-seat SaaS, building systems that show their work, and getting AI into production. Short, specific, no hype.

Shipping real ML

June 13, 2026 · 8 min read

The ML was the easy part

A weekend pipeline that isolates a vocal, swaps the singer with a Retrieval-based Voice Conversion model, and renders a promo video, fully automated. The hard part was not the model. It was making a fragile, half-abandoned dependency stack run reliably across a Windows box and a Colab GPU.

Read the post →

The agentic build loop

June 8, 2026 · 5 min read

A content pipeline that can't go off-brand

I do not trust an AI to remember the legal disclaimer, so I do not ask it to. Deterministic code enforces the non-negotiables before any reviewer reads the draft. That is what makes unattended publishing safe to leave alone.

Read the post →

June 8, 2026 · 5 min read

I replaced a scoring rubric with a tournament of identical judges

I deleted a hand-weighted scorer and ran three randomized brackets of identical LLM judges instead. The disagreement between them turned out to be a better confidence signal than any weight I set by hand.

Read the post →

June 6, 2026 · 5 min read

Who reviews the AI's code? An independent reviewer with a fixed rubric

When an AI agent writes most of the code, self-review is worthless. Here is the independent reviewer, the fixed six-category rubric, and the measured ledger that keep it honest.

Read the post →

June 6, 2026 · 5 min read

Architecture that defends itself: encoding decisions as lint checks

Architecture decisions in a wiki erode. Encoded as custom lint checks that fail the build, they enforce themselves, even against the next AI-generated diff.

Read the post →

June 6, 2026 · 5 min read

How one engineer ships production software on a real SDLC

No QA team, no release manager, several deploys a week. The lifecycle is not discipline I remember to follow. It is encoded as commands an AI coding agent cannot skip.

Read the post →

Own the software you run

June 8, 2026 · 5 min read

One person runs a whole daily operation

A daily operation that would conventionally need several people runs as a dozen small scheduled programs with one person supervising. The bottleneck was never headcount. It was nobody designing the workflow as software.

Read the post →

June 4, 2026 · 5 min read

Replace per-seat SaaS with software you own

Per-seat ERP and BI licenses scale with headcount, not value. Here is how I replace them with custom apps that run in your cloud, and what one client saved.

Read the post →

Grounded, trustworthy AI

June 8, 2026 · 6 min read

Your data was lying long before the model was

I scrutinized every word the model produced and trusted the numbers my own code fed it. An adversarial pass found thirteen silent data bugs, including fabricated zero-width spreads on roughly 43 percent of candidates.

Read the post →

June 6, 2026 · 5 min read

Give the AI agent a map: a provenance-tagged knowledge base for a legacy system

Point an AI agent at a legacy system cold and it guesses fluently. Here is the compiled, provenance-tagged knowledge base that lets it work from the truth instead.

Read the post →

June 6, 2026 · 5 min read

A 200 is not proof: verifying every write to a system of record

The most expensive bug I shipped returned a 200 and never landed. Here is the verify-after-write pattern, and the lint gate that makes skipping it impossible.

Read the post →

May 28, 2026 · 6 min read

RAG that cites every source: evidence search for legal text

Retrieval-augmented generation is useless in law if it cannot show its work. Here is how I built text-message evidence search where every result traces to a timestamped source.

Read the post →

Agents in production

June 6, 2026 · 5 min read

Permission keys, not roles: authorization that survives real org complexity

Role-based access control collapses into special cases the moment an org gets real. Here is the permission-key engine, one registry, audit log, and delegation that replaced it.

Read the post →

June 6, 2026 · 5 min read

The agent that operates the cloud, not just writes the code

The real value of an AI agent is not typing code. It is driving the cloud from the CLI: opening PRs that merge themselves and pinning the right config to the right slot.

Read the post →

June 6, 2026 · 6 min read

The deploy I never think about: zero-downtime releases that roll themselves back

Two outages taught me to stop hoping. Build-stamped smoke gates, run-from-package, and atomic slot swaps mean you can never silently ship the wrong build.

Read the post →

Have a problem worth solving?

Thirty minutes, no scripts. Describe the work and I'll tell you straight if I can help.

See services and pricing Book a 30-minute call