What I've shipped.
The AI systems I've designed and shipped to production: MCP infrastructure, architecture assistant, anomaly detection, RAG, observability. The problem, what I built, and what it changes.
Case studies
The problem, then the result.
Every project starts from a concrete need. Here's what I built, and what it changed.
- AI infrastructure
- Security
Give AI agents access to internal tools without opening a hole.
The shift
Every LLM wired to each internal tool by hand
Teams wire an AI onto an internal tool without opening extra access
What I built
An MCP infrastructure that decouples LLMs from executing information-system tools: a Zero Trust / SSO proxy in front of every tool, secrets never exposed, and skills that encode how to use the log platform well.
The stack
MCP · Skills · fastmcp · Zero Trust · SSO
- Agents
- Human sign-off
Have an agent write the architecture record it cannot confirm.
The shift
Records retyped by hand from one deliverable to the next
An entry nobody has confirmed reaches no report and no client link
What I built
An architecture assistant wired into the project record: it writes typed entries across twelve registers, and it holds no tool to confirm them. Human sign-off is not a rule anyone can forget, it is a missing service.
The stack
LangGraph · FastAPI · PostgreSQL · React · TOGAF
- Anomaly detection
- Cost
Price the capacity paid for nothing, with no model inventing a number.
The shift
Capacity thresholds watched by hand, and nothing priced
The capacity gap is priced by month and year, with the sum shown
What I built
A detection chain that is fully deterministic: peer cohorts, robust statistics, a control that stops the volume of tests from making alerts. The model only rephrases: if it adds a number, the computed text is what ships.
The stack
Python · FastAPI · scikit-learn · Next.js · LLM
- RAG
- Knowledge base
A technical knowledge base that answers, always in sync with the code.
The shift
Docs drifting from the code, answers hunted down by hand
Teams get a sourced answer instead of digging through stale docs
What I built
An end-to-end RAG chain feeding the internal AI assistant, and an idempotent synchronization service between the technical GitHub repositories and the knowledge base.
The stack
RAG · Elasticsearch · Kibana · Python · GitHub
- Observability
- LLM
Catch the anomaly the moment the log is ingested.
The shift
Anomalies lost in the volume, caught after the fact
The anomaly is flagged the moment it lands, not after the incident
What I built
A native Logstash filter, in Ruby, that calls an LLM during log ingestion and returns anomalies as structured JSON, ready to use.
The stack
Logstash · Ruby · LLM · Elastic Stack
Your project here
The next one could be yours. Tell me what you want to build.
Work done on assignment. Full context on request.
Contact
Ready to go from demo to production?
Reply within 24 hours · first conversation free, no strings attached.