The platform every other AI project stands on: model gateway, retrieval, caching, cost control, evaluation and observability, running somewhere you control.
Services
What I do
Six ways to work together. They are deliberately layered: infrastructure first, then agents, then the customer-facing surfaces — because that is the order in which AI projects actually succeed or fail.
Most companies arrive with one of three problems: they have an idea and no platform, they have a pilot that collapsed under real load, or they have a working demo that nobody can put into production for legal reasons. Each service below is built for one of those situations.
You can engage a single service or a sequence. If you are not sure which one you need, the FAQ is the fastest way to work it out, or just ask me.
Deliverables
Multi-step systems that act inside your tools — CRM, ERP, ticketing, documents — with permissions, budgets, audit trails and human approval where it counts.
Assistants grounded in your real content that answer, qualify, hand off to a human and report what they actually resolved.
The retrieval layer: chunking, embeddings, hybrid search, reranking, freshness and citations — so answers stay correct when the documents change.
Golden datasets, LLM-as-judge, regression suites and dashboards for quality, latency and spend — so you can change models without gambling.
The non-AI half: APIs, integrations, internal tools, IoT and web platforms, built to outlast the current framework trend.
How engagements usually run
Audit (1–2 weeks)
I read your code, data and constraints and talk to the people who will live with the result. You get a written plan: what to build, what to skip, what it costs, what can go wrong.
Foundation (2–4 weeks)
Platform and retrieval work: gateway, data pipeline, evaluation harness, deployment. Small, boring, reviewable — and the part that makes everything later safe.
Ship (4–8 weeks)
The actual user-facing capability: an agent that closes a queue, or a website assistant that answers in three languages. Measured against a baseline agreed in week one.
Operate (ongoing)
Monitoring, prompt and retrieval tuning, model migrations, incident response. Optional — most clients keep me on for a few days a month and that is when the value compounds.
What changes when the work is done
- AI features stop being demos and become part of the release train, with tests and rollback.
- Answers are traceable to a source document, so legal and compliance stop being the bottleneck.
- Per-feature model cost is visible, and a model swap is a config change rather than a project.
- Your team can debug an AI failure the way they debug any other bug — from a log, not from a hunch.
- The system runs on infrastructure you control, with a provider exit that does not require a rewrite.
Typical stack
- TypeScript
- Node.js
- Python
- Claude / GPT / Gemini
- MCP
- pgvector / Qdrant
- PostgreSQL
- Redis
- AWS
- Docker
- Kubernetes
- OpenTelemetry
Frequently asked questions
Do I need all six services?
No. Most engagements use two or three. A typical sequence is AI infrastructure, then the specific surface — agents or a website chatbot. RAG and evaluation are usually pulled in as part of the foundation phase, because they are what make the rest survive production.
Which model providers do you work with?
Anthropic, OpenAI and Google are the default, but I design for provider independence: one gateway interface, embeddings and retrieval behind it, so switching or mixing models is a configuration change. If you need a specific model for data residency reasons, that is fine too.
Can you work with our existing stack?
Yes, and I prefer to. If your services are Python or Java, the AI layer will be too — I will not rewrite working infrastructure to fit my preferences. The only non-negotiable is that the parts touching your data are testable.
Do you sign NDAs and work under security review?
Regularly. Enterprise AI projects usually require a mutual NDA before anyone sees a system diagram, and I am comfortable going through your security questionnaire. I can also work in a fully local environment if your data cannot leave the network.
What does it cost?
See the pricing page for models and ranges. The short version: an audit is a fixed fee, delivery is day-rate, and I quote a full phase estimate before starting — not a discovery that turns into an invoice.
Related writing
-
How to Introduce AI Agents Without Burning the IT Budget
A practical rollout plan for enterprise AI agents: pick one queue, run in shadow mode, draft before acting, and widen autonomy only where the numbers hold up.
-
RAG or Fine-Tuning? Choosing the Right Approach for Knowledge
When retrieval-augmented generation beats fine-tuning — and why almost every company knowledge problem is really a retrieval problem wearing a disguise.
-
What an AI Chatbot on a Website Actually Returns
How to measure a website AI chatbot: deflection, resolution, unanswered-question reporting and the four metrics that predict whether it is worth keeping.