AI assurance services for production AI

What the engagement answers

Most teams can show a demo. Fewer can answer five questions without calling the vendor:

• Does it work repeatedly? Not once, not on a clean test set — on real tasks, at k ≥ 3, with route and tool-call variation.
• What does a successful outcome cost? Tokens, retrieval, reranking, retries, and human review, measured per accepted result.
• How does it fail? Which inputs, tools, permissions, and edge cases produce a wrong or unsafe outcome?
• Can we reconstruct it? Are traces, model versions, retrieved evidence, and approvals retained?
• Would this survive an audit? Is there a control matrix, an owner, a severity-ranked findings register, and a remediation plan?

What you get

• A system and data-flow map with the assumptions that matter.
• A scenario set derived from real workflows, support tickets, and incidents.
• A k ≥ 3 evaluation with route, tool-call, latency, and cost traces.
• A severity-ranked findings register with owners.
• QA/evaluation procedures and a control matrix mapped to the framework you actually need.
• A regression harness installed in your CI/CD, owned by your team.
• A 10–20 page report, board deck, and 60-minute readout.

Engagements

Engagement	Time	Price
AI Readiness Diagnostic	1 week	$7,500
Full Production Assurance	2 weeks	$25,000
Fractional Assurance Officer	20 h/week, 90-day minimum	$20,000/month

The Diagnostic is credited 100% toward a larger engagement within 60 days. See the full engagement ladder and scope.

Who this is for

US and Canadian teams with an AI agent, RAG system, or model already in production or about to ship. Best fit: a named engineering owner, access to traces and evaluation data, and a release, audit, board, or customer deadline. Regulated fintech, health, insurance, and enterprise platform teams are the strongest fit.

Why independent

I do not resell an evaluation vendor, an observability platform, or a governance SaaS. The recommendation is based on what the system needs, and the harness is handed to your team. The evidence has to stand on its own after the engagement ends.

The work is led by a Principal AI Assurance & Information Systems Specialist with an MS in Information Systems from Northeastern University and an NVIDIA NCA-GENL credential. Public proof includes KYC Copilot (own product, open-sourced) and an anonymized nonprofit data system.

How the work runs

1. Scope: one system, the axes in scope, access, success criteria, and the readout date.
2. Assess: trace review, stakeholder interviews, a golden set, and k ≥ 3 runs against the real workflow.
3. Readout: a 10–20 page report, board deck, severity-ranked findings, and a 30-day plan.

No code changes happen before the SOW is signed. Every check has an evidence link; every P0 has an owner.

What this is not

It is not a certificate, a generic code review, staff augmentation, or a resold observability platform. It is an independent engineering assessment that produces evidence your team can run and defend after the engagement.

What success looks like

Success is not a prettier dashboard. It is a signed-off evidence pack, a regression gate your team can run, a severity-ranked remediation list with owners, and an executive who can explain what was tested and what remains open. If the system is not ready, the report says so plainly.

US and Canada delivery

Remote-first with ET/PT overlap. Invoices in USD; USD and CAD settlement through Wise; Deel contracts where required. Cal.com displays your local time.