SPECIMEN №01 · AI BUSINESS ARCHITECT
FIGHTER PILOT · REGULATED INDUSTRIES · ONE ACCOUNTABLE BUILDER
BY THE NUMBERS
€5M/yr
Manual process retired
DocFields · projected 12-mo run-rate
100+ FTE
Equivalent workload
Purchase validation · field UAT
1+ yr
In production
Onboarding engine · 35 nodes · 3 languages
<1s
Voice catalog response
Askalog · 24,000 SKUs · zero hallucinated prices
8
Products shipped solo
4 prod · 2 pre-prod · 1 alpha · 1 build
ABOUT
Your company is paying people to read documents, copy fields, and chase each other for approvals. I find that process, I put it in front of its own numbers, and I retire it: specified first, built end-to-end, in production with real users in weeks. €5M a year retired so far, 100 people worth of work, in one regulated group.
One person means one accountable. You get the same builder from the first call to the first production incident. And you never depend on me: every system ships with the spec it was built from, every decision written down, tests on every delivery, and a handover your team can run without me.
Operational discipline from a decade in a cockpit: understand the operation before touching it, no improvisation on critical fields, no undeclared exceptions in production.
WHAT I'VE SHIPPED, ALONE, TO REAL USERS
A document intelligence platform with four extraction engines and contract generation. It retired a €5M/year manual process (projected, 12-month run-rate) and the equivalent workload of 100+ people. Production for over a year.
A client portal that walks the seller of a vehicle through 35 decision nodes and 34 document types in three languages, and hands a clean, cross-checked file to the back office. Production, one year.
Autonomous validation of purchase files in an enterprise back office: every field checked against every other document, field-level provenance, permissions by job. Nine findings from a field review closed in 48 hours.
A business-native AI assistant that lives inside the company's own system: it reads the case files, the documents and the law it must comply with, and acts through the platform's tools, gated per tenant and traced end to end.
Prototype-to-production cadence: a working demo in days, an integrated system in weeks, and a written trail a CTO can audit.
I've done this in automotive finance, aviation, fintech, legal and defense. The method is what transfers: learn the operation faster than the people who run it expect, challenge the requirements they accepted as fixed, then use AI to build past them.
Strategy, execution, and the on-call phone number. One person, accountable.
SELECTED WORK
Four in production. Two pre-production. One in alpha. One in build.
Situation: a regulated group extracting fields from 34 document types by hand. Result: €5M/year retired, one normalized output, live in production.
Twenty-four enterprise processors, confidence-tiered fallback, one normalized output. Most contracts stop at the cheapest tier.
Client portal for an auto retail group · powered by DocFields.ai
Cross-document conflict detection, append-only attestation per field. No human keystroke between document upload and the underwriting payload.
Decision-tree document-set engine · live for over a year
Thirty-five nodes. Thirty-four document types. Three languages. The graph decides which papers a customer must upload — and which stack to demand when the easy answer fails.
Situation: 100+ people re-reading purchase files. Result: files clear on their own; operators see only what disagrees. Field QA: nine findings, closed in 48 hours.
Every field on the contract cross-checked against every other document in the bundle. Most files clear without an operator touch.
Situation: a multi-tenant professional-services platform with an AI that knew nothing about its own cases. Result: an assistant that lives inside the platform, reads the files, and acts through its tools.
It knows the case, the document, the calendar and the law that applies, because it runs inside the system that holds them. Every capability gated per tenant. Every action traced.
Sub-second voice catalog agent · askalog.com (alpha)
Ask in Spanish, get the car. Sub-second, deterministic, zero hallucinated prices — the unit economics of voice commerce, finally.
Real-time agentic translation for regulated meetings
KUDO sells you interpreters. Wordly sells you captions. SmithVox sells you a tamper-evident multilingual transcript — in the speaker's own voice — and a chair agent that keeps the meeting from drifting.
Aspect-based sentiment with sector RAG · whyrating.com (pre-production)
Thirty-six primitives, intensity-scaled, anchor-grounded to the verbatim review text. Reputation noise becomes a queryable underwriting signal.
CURRENTLY SHIPPING
Week of Sep 22, 2026
A field review by the operations team raised nine findings. All nine closed in 48 hours: one URL per file, permissions by job, validated identity written back to the ledger.
LLM escalation for fuzzy matches now ships with an explicit degraded flag. The consumer discards degraded verdicts instead of trusting them.
Per-tenant Google Workspace connectors and legal RAG tools behind tenant flags. Every tool call traced.
Updated by hand, not by a job.
THE FIRST ONE TO GO
Somewhere in your company a process still runs on people re-reading documents. That is the one I retire first.
HOW WE WORK
Anchored on outcomes, not hours. Every step is credited against the next, and every delivery ships with spec, tests and handover.
01 · DIAGNOSTIC
I map the process, measure what it costs you today, design the system that retires it, and write the plan. You keep the document whether or not we continue.
Fixed fee · credited against the pilot
02 · PILOT
End to end: data model, extraction, rules, interface, integration with what you already run. Specified before it is built; tested on every delivery; handed over with the decisions written down.
Half on start · half when it is in users' hands
03 · RETAINER
Same builder, same method, a standing capacity to keep retiring manual work and to answer the phone when production calls. Priority on your roadmap, no re-onboarding.
Monthly · cancel with a month's notice
If the diagnostic finds no case worth building, I say so, and you keep the report.
METHOD
Learn the operation faster than the people who run it expect, challenge the requirements they accepted as fixed, then use AI to build past them.
PHOTO №01 · F/A-18 HORNET
BACKGROUND
A decade as a fighter pilot and Operational Safety Officer in the Spanish Air Force. Operational discipline transferred to software architecture: systems either work under pressure or they fail people. That standard runs through every project here — the audit trails, the deterministic guards, the field-level provenance.
The same loop in both jobs: understand the operation first, then build what it actually needs. No improvisation on critical fields. No undeclared exceptions in production.
CONTACT
If you are a CTO, head of operations or founder in a regulated business, book the diagnostic call. You leave with three things: which process to retire first, what it costs you today, and whether I am the right person to do it. If there is no case, I say so.
Address revealed on click · spoofing-resistant
ALSO BUILDING
FRONTBUG.COM ↗
Never explain a bug twice.
Right-click what is broken; the report writes itself, ready for your AI agent. QA, smoke tests and documentation, in one place.
WHYRATING.COM ↗
Reputation, decoded.
Aspect-based review intelligence. Powered by Prism Engine.
CLAUDEMESH.COM ↗
Peer mesh for Claude Code.
Connect agent sessions. End-to-end encrypted. Zero config.
