Anyone can wire a model API in a weekend. What kills the feature is everything after: retrieval that is wrong on your data, quality that drifts, a security review it cannot survive, a cost per request nobody measured, and a system only one person understands. I have hit all five in production — at enterprise scale, on someone else’s budget.
Not the model. Every one of these is an architecture problem, and every one has a receipt behind it.
Rates depend on the shape of your data and the depth of the guardrails. I quote after I have seen both — never before.
Retrieval quality against a real question set, an eval baseline your team can keep running, the guardrail gaps that will surface in diligence, a measured cost per request, and a 90-day plan. Delivered written, so it is useful whether or not we work together after it.
Data, retrieval, orchestration, evals, guardrails, deploy. It ships behind a flag with a golden set, a runbook and a cost model — not a notebook and a demo video. Scoped to one workflow that has a real user waiting on it, because that is the only scope that finishes.
Architecture and design reviews, a reusable guardrail kit your engineers apply across features, eval discipline that outlasts me, and help setting the hiring bar for AI engineers. For teams with good engineers and nobody who has taken an LLM system to production before.
For companies where AI is the product rather than a feature, and the work means sitting with customers where the system meets the real workflow. Open to remote Forward Deployed Engineer and Principal / Enterprise / Senior AI Architect roles.
The order matters more than the speed. Nothing gets built until we agree, in writing, what a correct answer looks like.
Enterprise work sits behind an NDA, and you are right to discount it. These you can open, read and pick apart without talking to me.
Four bounded agents under a supervisor that is deliberately not an agent, a citation guard checked in C#, fourteen deterministic invariants and a hash-chained ledger. The design line: the model reads, retrieves and explains — it never sets the monetary verdict. Production-shaped, not production-deployed, on synthetic data — and the write-up says so.
Read the case study →The generalised architecture behind the framework running 20+ production AI workflows — hybrid retrieval that survives identifier queries, entitlement filtering pushed inside the query, a semantic cache that knows when it is lying, and why a promotion peak is the honest test of shared AI infrastructure.
Read the case study →Five narrow agents on .NET 10 and Semantic Kernel, including a hallucination guard that validates every output against the patient’s own FHIR record. Consent, PHI redaction, tool allow-lists and an audit trail. Read the repository before you read my résumé.
Open the walkthrough →Enterprise ERP, healthcare, education, and offline-first designs for low-resource settings. Click any node for its responsibilities, tech and standards, controls and NFRs, a worked example, and what happens when it fails. Except the HealthQCopilot walkthrough, these are reference designs rather than shipped products.
Browse the architectures →Thirty minutes. Bring the workflow you would most like to automate and, if you have one, an example of it going wrong. You will leave with an honest read on whether it is a retrieval problem, an eval problem, a guardrail problem — or not an AI problem at all. Whether or not you hire me. If it is not a fit, I will say so on that call and tell you what I would do instead.