What the engagement covers
The model is the easy part. The platform under it is the work
Most AI projects do not stall on model quality. They stall on retrieval that returns the wrong chunk, a GPU bill nobody forecast, an evaluation loop that does not exist, and no answer for what happens when the model says something it should not. That is infrastructure work, and it is the same discipline as the rest of this site.
RAG architecture
Chunking, embeddings, hybrid retrieval, re-ranking and caching, designed around your corpus rather than a tutorial. Vector search on Azure AI Search, pgvector, Pinecone or Chroma.
Guardrails and evaluation
Prompt-injection mitigation, output guardrails, and an evaluation set that runs in CI, so a prompt change is a tested change rather than a hopeful one.
LLM observability and cost
Token and spend tracking per team and per developer, with latency and failure tracing through OpenTelemetry. The FinOps problem moved to AI workloads, which is where it is least governed.
Model serving and GPU scheduling
NVIDIA operators and Triton on Kubernetes, with scheduling and autoscaling that reflect what inference actually costs.
Agentic and tool-calling systems
Tool and function calling, MCP integrations and agent workflows, given the same treatment of secrets, identity and blast radius as any other production system.
MLOps CI/CD
Azure ML, Vertex AI and SageMaker pipelines wired into the delivery model you already run, rather than a parallel one only the data team understands.
- Claude
- OpenAI
- GitHub Copilot
- NVIDIA
- Azure OpenAI
- Kubernetes
Where this shows up in the work
The AI work is the newest layer, and most of it lives in Alertloop — the observability product I build and run.
Common questions
Do you build models, or the platform around them?
The platform. Retrieval, serving, evaluation, observability, cost and security. If you need a novel model trained, that is a different specialism — but almost nobody needs that, and almost everybody needs the platform.
Our AI spend is unpredictable. Can you help?
That is the same problem as cloud billing, one layer up, and it is where I go deepest. Token and spend tracking per developer and per feature, then the same rightsizing and commitment logic applied to inference.
Is a RAG system a security risk?
It can be, and the risks are specific: prompt injection through retrieved content, over-broad retrieval exposing documents a user should not see, and secrets reaching a model context. Each has a known mitigation, and they belong in the design rather than in a later review.
Can this run inside our compliance boundary?
Yes. Azure OpenAI inside your own tenancy, self-hosted models on GPU nodes in your own cluster, or a hybrid. Most of my regulated work predates the AI layer, and the constraints are the same ones.
Other things I'm brought in for
Most engagements touch more than one of these.
Bring me the AI system that costs more than it should.
Thirty minutes is enough to find out whether the problem is retrieval, serving, evaluation, or simply nobody watching the token bill.