Make your RAG or agent safe, observable, and cost-bounded in production.
I review and fix data access, tool permissions, prompt-injection exposure, retention, observability, failure handling, and cost controls for live AI features.
For live LLM, RAG, and agent systems — not idea-stage AI productsThe control surfaces that decide whether an AI feature is production-ready
A demo proves the model can answer. Production readiness is everything around it.
- Data access: what the model and retrieval layer can actually read, per user and per tenant
- Tool permissions: what an agent is allowed to do, with what credentials, and what stops it
- Prompt-injection exposure across retrieved documents, user input, and tool output
- Retention: what prompts, outputs, and tool calls are stored, for how long, and where
- Observability: logging and tracing that make failures and abuse diagnosable
- Failure handling: timeouts, fallbacks, and degraded modes instead of silent wrong answers
- Cost controls: spend caps, rate limits, and per-tenant budgets that hold under real traffic
Public tooling and technical notes you can inspect first
Open-source work and educational material — labelled for what it is. Representative demonstrations are worked examples, not client case studies.
LLM production-safety scanner
An open-source CLI for testing production safety controls in LLM and RAG applications — prompt injection, data leakage, and cost vulnerabilities. Public, MIT licensed.
View the repository →AI workload control notes
Technical articles on prompt injection, PII leakage, agent tool-use security, LLM observability, cost controls, and safety regression testing.
Browse the resources →Is this the right engagement?
A quick self-check before the call.
Good fit
- A live or launching LLM, RAG, or agent feature inside a real product.
- A named concern: data exposure, tool permissions, prompt injection, runaway cost, or undiagnosable failures.
- A team that can grant repository and configuration access for review and fixes.
- A buyer, security review, or incident creating real pressure to fix it now.
Not a fit
- An idea-stage AI product with nothing running.
- A request for model training, fine-tuning, or ML research.
- A formal AI-safety audit, certification, or legal opinion.
- General AI strategy consulting detached from a concrete running system.
Common questions
What does an AI production readiness engagement look like?
A review of the running system against the control surfaces above — data access, tool permissions, injection exposure, retention, observability, failure handling, and cost — followed by fixing the agreed gaps in your codebase and infrastructure. Findings come with a fix, not just a flag.
Can you help with a RAG security review a customer or auditor is asking for?
Yes — engineering evidence for the controls that exist and implementation of the ones that are missing. This is engineering remediation and evidence support, not a formal attestation or penetration test.
Our LLM costs are unpredictable. Is that in scope?
Yes. Cost controls — spend caps, rate limits, token budgets, caching decisions, and per-tenant metering — are part of production readiness, and unbounded spend is one of the most common gaps in shipped AI features.
How do you test for prompt injection?
By examining every path where untrusted text reaches the model — retrieved documents, user input, tool output — and testing the controls around them. I maintain an open-source scanner for exactly this class of checks, which you can inspect before we talk.
Is AI observability different from normal application monitoring?
It adds model-specific signals: prompt and tool-call logging with retention rules, token and cost metrics, quality regressions, and abuse patterns. The engagement wires these into the monitoring you already have rather than introducing a parallel stack.
Tell me what the AI feature does and what worries you about it
Bring the system, its users, and the concern — data access, injection, cost, reliability, or a security review asking questions you cannot yet answer.