Auto Pilot Events OS
An AI operating system for an established UAE experiential events group operating across the GCC. An institutional brain that ingests everything the business produces, learns its economics, and runs the deal lifecycle from RFP to production under human supervision. Architected, built and shipped in 28 days by directing a fleet of AI coding agents.
The problem: every events company runs on tribal knowledge
I have spent twelve years inside large-scale live events in the UAE: GITEX, COP28, Dubai World Cup, government galas. The pattern is the same everywhere: the company's real intelligence lives in a handful of heads and a graveyard of old Excel sheets.
An RFP lands. Someone digs through past quotations to remember what a 6-metre stand, an LED wall or a custom counter actually costs. Pricing is inconsistent between estimators. Supplier knowledge is personal, not institutional. When a senior estimator leaves, years of pricing instinct walk out the door with them.
Nothing the company learns on one deal makes the next deal any smarter.
That is the actual transformation problem, and it is not "add a chatbot." The fix is structural: a system of record that turns every document, every quote and every correction into compounding institutional memory.
The design bet: five principles before a line of code
I wrote a full 360° blueprint before building, anchored on principles chosen to survive to 2040, not to demo well next quarter:
The default operator is an AI agent; humans supervise through approval gates. The UI is a window into what agents are doing, not a pile of forms.
Pricing, 3D-file parsing, BOQ splitting, supplier discovery. Each is one typed, documented tool. A human clicks a button; an agent calls the same tool. No human-only capabilities.
An append-only log of every business event is simultaneously the audit trail, the agents' working memory, and the ML training data. One design, three payoffs.
If a paid AI engine is unavailable, every feature still runs on a cheaper or local one. Nothing crashes, nothing blocks. This is what makes the $0 mode real.
A model ships only when it beats the working heuristic on cross-validated evidence. No model zoo you cannot feed. That is theater.
What was built
- The deal lifecycle, end to end: RFP intake and analysis, 3D design-file parsing, pricing suggestions with confidence tiers, commercial drafting in the house style, supplier RFQ splitting and quote comparison, production planning, scheduling and resource booking.
- 95 registered tools across 29 families behind one interface: pricing, deals, sourcing, commercial, finance, scheduling and the rest. Every capability callable identically by a human clicking a button or an agent doing its job.
- 7 specialist agents plus a director: intake clerk, pricing analyst, sourcing scout, commercial drafter, render reader, QA verifier, data scientist, working the event log around the clock, plus a natural-language assistant for the whole system.
- A 23-page operations web app: dashboard, deals, pricing, sourcing, commercial, finance, analytics, approvals, audited down to 375-pixel phone screens.
- Institutional memory that compounds: 1,136 price points learned from the company's real historical documents at ship date, growing with every deal ingested.
The pricing brain: measured, not promised
The heart of the system learns what things cost from the company's own history. All results below were measured on real historical pricing data, cross-validated, and verified live against the running system, not projected.
The part most AI projects skip: the first trained model looked 31% better than the regression, but only 1% better than a simple median. The system's promotion gate rejected it as theater. It also demotes a previously live model whose retrain no longer clears the bar, and when asked to break down a cost section it has too little history for, it refuses and says so rather than inventing a split. Honesty is enforced by machinery, not by intention.
Governance: autonomy is earned one capability at a time
Directors don't get asked "can you build AI?" They get asked "can you control it?" Every agent action passes through safety tiers, and each capability climbs this ramp only as evidence accumulates:
Drafts only. A human executes everything.
Acts, but a human approves each action.
Self-approves low-risk actions within rules; asks on high-risk.
Runs routine deals end-to-end; escalates exceptions only.
- Hard-blocked tier: sending money, sending external messages, deleting data: blocked until a human explicitly lifts them, regardless of how capable the agents become.
- Full journal: every agent step is recorded on the append-only spine. The audit trail isn't a feature; it's the architecture.
- Evaluation harness: a "smarter" prompt or model that breaks pricing fails the test suite before it ships. Stronger must be measured, not claimed.
Cost engineering: four modes, zero breakage
One environment variable moves the entire system between four cost postures: same containers, same database schema, same features everywhere. Capabilities degrade gracefully; nothing disappears. An events company can start at literally zero and pay only when volume justifies it.
| Mode | Runs on | AI engines | Cost / month |
|---|---|---|---|
| Local free | One laptop + Docker | Free cloud + local models | $0 |
| Cloud free | Free-tier cloud server | Free cloud models | ~$1 |
| Boosted | Same server | + paid frontier models | ~$15–40 |
| All-in | Scaled cloud | + paid frontier models | ~$100–300 |
How one person shipped this in 28 days
I built this the way I'd run an AI-transformed company: as the director of a fleet of AI coding agents, not as a lone typist. Frontier models drew the architecture and reviewed the hard decisions; workhorse models implemented scoped, pre-specified batches; every batch shipped with tests and was verified live against the real system before the next began.
- ~39,000 lines of working code: a 25k-line Python backend across 222 files and a 14k-line TypeScript frontend, plus 8,800 lines of tests.
- 753 automated tests, all green: grown batch-by-batch, never bolted on after (re-run and verified the day this case study was written). The suite is what lets AI agents build safely at this speed.
- Live verification as a discipline: measured claims (the pricing numbers above) come from runs against the real deployed system, and the build log records the failures too, including the model the promotion gate rejected.
- Scope discipline: a deferred-work ledger tracks every consciously parked feature with the condition that revives it. The build queue was closed by decision, not exhaustion: the system now improves by being used.
That workflow, blueprint first, typed tool seams, statistical promotion gates, human approval tiers, test-gated AI batches, is the same operating discipline I bring to transforming a business. The codebase is the proof it works.
What I will not claim
- It is not finished learning, and that is the design. The hard work is done: the corpus is built, the pipeline runs, the promotion gate works. Accuracy now improves with volume, so the more history a company feeds it, the sharper it gets. Any vendor telling you their pricing model is finished on day one is selling you something.
- No revenue-ROI figures yet. The system entered production use on July 12, 2026. Client-facing ROI numbers (hours saved per quote, win-rate movement) will be published once real usage produces them, measured like everything above.
- The accuracy numbers are technical, not commercial. They are cross-validated error reductions on real historical data. Strong evidence the learning works, not a promise about your P&L.
- Some capabilities are deliberately dormant. Deal-outcome prediction, for example, activates itself only after ~100 closed deals of data exist. It would be theater to switch it on sooner, so it's off.
If a consultant shows you an AI case study with no section like this one, ask why.