Purple Orange AI
Short version: The talk shows how Violema, a production multi-tenant SaaS, gets built by one founder without an engineering background: an agent swarm does the work, and reviewers with real authority decide whether it ships. The August 11 main-stage slot was cancelled on the day; the demo is re-submitted for September 15.
Agent swarms make code cheap. Parallel build lanes produce in a day what a small team used to ship in a sprint. The failure mode arrives just as fast: regressions ship at the same speed as features. Velocity without a gate is how a one-person SaaS dies in production.
The talk is about the gate. Every Violema deploy passes two adversarial reviewers, Claude Fable 5.1 and OpenAI GPT-6 Astra, and the stricter verdict binds. When the work is not ready, the verdict is DO-NOT-SHIP and the lane goes back to work. The logs of those refusals are the most useful artifact the system produces, and they are part of the demo.
Violema itself is live and came through a controlled beta built around the same principle: drafts, approvals, and an inspectable ledger before anything becomes final.
The architecture
Build lanes
Parallel Claude lanes under an orchestrator. Each lane gets narrow scope and a return contract: files changed, test results, a short summary.
Adversarial reviewer
A separate model reads the diff, the tests, and the mission objective. Its verdict is binding. DO-NOT-SHIP sends the work back.
File-based memory
Context survives across sessions, so lanes resume work instead of rediscovering it.
Approval gates
Deploys, migrations, and anything customer-facing wait for a person.
A reviewer that can only comment is decoration. The gate works because it can stop the line.
What the demo shows
Five minutes, live, no slides:
- Lane orchestration running on real work, not a staged example.
- Actual DO-NOT-SHIP review logs: the deploys the reviewer blocked and the regressions it caught.
- One full mission inside Violema: source tracking, approval gates, and the cost ledger.
- The backend test suite growing from 213 to 448 tests in three days, and what that did for reviewer confidence.
Why the gate matters for client work
The same pattern runs through every Purple Orange AI build: agents draft, prepare, and execute, and a control layer decides what becomes final. Separate the models. Separate the incentives. Give the check the power to stop the line.
If an AI workflow can write to production systems and nothing in the pipeline can refuse the write, you do not have automation. You have exposure.
That is why review gates, approvals, and telemetry show up in every workflow we ship, from a free workflow audit through a full operations buildout. The talk shows the pattern at its most extreme: one founder, one production SaaS, and a reviewer with the authority to say no.
What changed since August
August 16. The first apples-to-apples run. Claude Fable 5 and OpenAI Sol reviewed the same nine commits with the same prompt. Fable said SHIP-WITH-FIXES, Sol said DO-NOT-SHIP, and each confirmed defects the other never saw. The commits stayed frozen until repairs passed both.
August 23. The post-deploy re-review refused the repair set: eight findings fixed, three partial, and nine follow-up findings, three of them defects the fix itself introduced, including a manual Run that no trial or Start workspace could use. The hotfix that followed is what runs in production.
September 8. Two frontier models released this month reviewed that hotfix with the same prompt. OpenAI GPT-6 Astra said DO-NOT-SHIP with seven findings. Claude Fable 5.1 said SHIP-WITH-FIXES with eight. Five clusters both found, two only Astra saw, three only Fable saw. A third model, Gemini 3.1 Pro run through Antigravity, was denied the repository by its sandbox and issued a confident verdict anyway: seven findings at line numbers that hold an import, two comments, and a blank line, with its own footer reading files read: zero. The verdict contract caught it in under a minute.
The rule that came out of it: the stricter verdict binds, repairs ship only on two greens, and every finding must name a file and a line that exists.
The event
AI Tinkerers Chicago is a hands-on meetup for builders shipping with LLMs. Demos are five minutes, live, code-first, no slides.
- Date: Tuesday, September 15, 2026. Demo re-submitted, acceptance pending.
- Time: 5:30 to 8:00 PM
- Location: Chicago. The address goes out to accepted attendees.
- Admission: selective. Registration asks for a link to your work: GitHub, LinkedIn, or X.
See it live
The demo is re-submitted for the September 15 meetup. Space is limited and attendees are screened, so bring a link to your work.
FAQ
When and where is the talk?
The August 11, 2026 main-stage slot at AI Tinkerers Chicago was cancelled on the day. The demo is re-submitted for the September 15, 2026 meetup, 5:30 to 8:00 PM in Chicago. This page carries the outcome either way.
What is the talk about?
How Violema, a production multi-tenant SaaS, is solo-founded with parallel Claude build lanes under an orchestrator and two adversarial reviewers, Claude Fable 5.1 and OpenAI GPT-6 Astra, with authority to block deploys. The stricter verdict binds. The demo is live and includes real DO-NOT-SHIP review logs.
What is Violema?
Violema is an outcome-first AI operator from Purple Orange AI for recurring founder and team workflows, with review gates, approvals, and an inspectable run and cost ledger. It is live at violema.com.
How do I attend?
Apply through AI Tinkerers Chicago for the September 15 meetup. Admission is selective and registration asks for a link to your work: GitHub, LinkedIn, or X.