For founders and CTOs whose AI initiative has stalled
The gap between an AI demo and an AI system is where projects die.
Every company can buy the same intelligence now. The models are a commodity; the advantage is where you point them and whether the thing survives contact with your business. That takes data plumbing, evals, guardrails, cost control and security. It is a production-systems problem, not a prompt problem.
That gap is my job: not advising from outside it, shipping from inside it.
Deployed forward
The work happens in your repo, your CI, your meetings. A keyboard, not a slide deck.
Production is the deliverable
Quality tests, tracing, guardrails and per-request cost budgets. They ship with the feature, not after it.
Leave-behind by design
Written decision records, runbooks, pairing. Your team owns a system, not a dependency on me.
01 · Engagements
Three ways to deploy me.
Every engagement is hands-on-keyboard, and every one ends with your team owning the result.
Embedded FDE Sprint
The core offering. One AI use case from whiteboard to production inside your team: scoping, data plumbing, evaluation harness, guardrails, ship. Fixed price from $15k, quoted after the recon call.
AI Solution Architecture
Design and the hard calls for your own build: model strategy, RAG and agent architecture, evaluation methodology, cost budgets. All of it lands as written decision records your engineers act on. Monthly retainer.
Production Hardening
Rescue the wobbling pilot: tracing, quality tests, guardrails, integration with the systems you already run (including mainframe-era ones), and a known cost per request. Scoped project.
Before you ask
Not sure which shape fits? The recon call decides. No cost, no pitch.
Book the recon callWhere I’m sharpest
The pilots that stall hardest are the ones that have to clear a security review.
02 · Field record
Twenty years of production, stated plainly.
Delivered at MuleSoft (Salesforce) · Platform architecture · 2016–2026
Designing a central LLM gateway for a government-serving SaaS platform: per-team quotas, cost projection, prompt-cache optimization, circuit breakers. CLIENT: ████████ (under NDA).
Freight-RFQ automation for a logistics operator: quote requests arrived as unstructured email; an LLM pipeline now parses them into structured, priced responses. Running in the business daily.
Agent-driven engineering workflow: scaffolding, migrations, tests, decision records. Human review on everything that ships.
A consultant tells you what to build. A forward deployed engineer builds it with you, in production.
03 · Approach
A deployment has five phases.
Anchored on written deliverables and working software, never on billable hours.
The 30-minute call
Plus a written read on where your biggest production risk lives. Fixed price quoted after; no surprises.
Inside your team
Least-privilege access to your repo and CI; data plumbing and scoping with your engineers.
Wired end to end
The use case runs against your real systems and your real data, on top of what you already have. No migration.
Non-determinism becomes evidence
A golden dataset of real queries with hand-labeled ideal outputs, then a measured pass rate, a named failure taxonomy, and defined escalation behavior. This is the phase that decides whether the system ships, and the one most pilots skip.
Harden & handoff
Guardrails, tracing, cost per request, runbooks, and written decision records. I roll off; your team owns it.
What you own on day one after I roll off
- The code, in your repo. Written with your engineers, in your stack, reviewed by your team.
- A golden dataset. Real queries with hand-labeled ideal outputs: the normal, the edge, the ambiguous and the high-risk ones.
- An eval harness and its pass rate. Re-runnable in your CI, so quality is a number that moves, not an opinion.
- A failure taxonomy. The named ways this system breaks: missing context, wrong tool, wrong record, invalid output, unsafe action, timeout. Each with defined behavior.
- Escalation rules. Where autonomy stops and a human takes over, decided deliberately rather than discovered in an incident.
- Cost per request. Measured, budgeted, and alerting before finance finds it.
- An audit trail. Every prompt, response, tool call, result and error, logged and queryable.
- Runbooks and written decision records. Why each choice was made, so the next engineer inherits reasoning and not just code.
Recon call to a production path: two to six weeks. Two when the use case is scoped and your data is reachable. Six when the integration surface is wide, the data needs work, or the deployment has to clear security and compliance review first. The recon call is where that gets decided, and the number is fixed in the quote before anything starts.
Self-assessment · 90 seconds · 10 checks
AI Production Readiness Score
Tick what your AI initiative already has. Honest answers only. The score is for you.
04 · About
The person you're giving prod access to.
Hablamos español. I work with US and LatAm teams in either language.
Dragging stubborn systems into the future since the mainframe era.
There is one way a system works and a thousand ways it fails. Walter Poch spent twenty years being paid for the thousand: tier-0 platforms where an outage was not a bad demo, it was every customer at once. Most recently as Software Engineering Architect at MuleSoft (Salesforce), owning the identity, authorization, auditing and cryptography layer that 250+ services depended on, leading the architecture team through a 250-microservice cloud migration, and running point as incident commander on customer-facing P0s.
That is the gap in most AI pilots. They are built for the path where the model answers correctly. Nobody owned the other thousand paths: the malformed response, the dead API, the timeout, the half-finished run, the confidently wrong answer that reaches a customer. Those failure modes are not new. They are the same ones distributed systems have always had, and they are what twenty years of tier-0 work teaches you to design for first.
He has been dragging stubborn systems into the future since he taught a 1980s Argentine bank mainframe to speak REST. Large language models are just the newest stubborn system.
Today he embeds inside your team and does the unglamorous work that gets AI from demo to daily production: data plumbing, automated quality tests, guardrails, and cost control.
Access & security: how embedding works
- Least-privilege access, in your repo and your cloud
- NDA by default
- Human review on everything that ships
- Decision records and runbooks are the leave-behind
- Your data never leaves your tenant
- References available on the recon call
Independent practice, established 2026. The record above is verifiable Salesforce/MuleSoft tenure.
Appendix: questions on file
Objections, answered up front.
We already have good engineers, so why you?+
Your engineers are good at your product. Production AI is a specialization that most teams meet for the first time mid-project: evaluation harnesses, guardrails, cost budgets, failure modes. I bring the patterns so your team doesn't learn them the expensive way, and the Architecture retainer exists exactly for teams who want to build it themselves.
What does the sprint cost and what do we own after?+
Sprints are fixed-price from $15k, quoted after the recon call. You own everything: the code in your repo, the evaluation harness, the runbooks, and the written decision records. There is no dependency on me when I roll off.
Our pilot is live but flaky. Is it too late?+
No. That is the most common starting point, and it is usually a two-to-four week job rather than a full build. Production Hardening exists for exactly this: tracing, quality tests, guardrails, and a known cost per request, wired into the systems you already run.
What do you need access to?+
Least-privilege access to your repo and CI, and only the systems the use case touches, under NDA by default. The Access & Security section covers the full arrangement; your data never leaves your tenant.
Why is one client unnamed?+
Because the work is under NDA and still in design. I would rather show you a redaction bar than inflate a claim. Every number on this page is anchored to work you can verify.
Do you work in Spanish, or with LatAm teams?+
Yes. Native Spanish, full-professional English. I'm US-based in South Florida and work with US and Latin American teams in either language and time zone.
05 · Open a channel
Tell me where the problem lives.
A few lines about your system and what's stuck. Personal reply within a day, EN or ES. Engagements from $15k.
The recon call is a deliverable, not a pitch. You leave with (1) a written read on why your pilot is stuck, (2) the production-readiness checklist, (3) a fixed-price plan if there's a fit.
If I'm not the right fit, I'll say so on the call and point you at who is.
Prefer email? me@wpoch.com