Enterprise AI architect · Context layers for AI agents · DACH
I help teams whose AI assistants already read chat, tickets, wiki and CRM answer from current, cited facts and take actions a person approves. I build the context layer on your own tenant, with delegated access, a temporal knowledge graph and an audit trail. We can start with one governed workflow; implementation is billed by the hour with an estimate for each phase.
Explore the work: 50-second demo and case study · Consulting approach · Discuss your workflow
The case study describes a production workflow. Its public demo and screenshots use fictional records to protect the client.
The parts of my client work that hold up in production become open source. Follow this profile to see the next ones when they ship, together with the implementation notes below.
- agentic-task-system gives an agent task and knowledge context over the tools you already use (CLI + MCP), with provenance, reviewed writes and undo. The underlying systems remain the source of truth.
- skillgate blocks commit or publish until your definition of done passes, for when a coding agent reports "done" too early. Works in Claude Code, opencode, pre-commit and CI.
- agent-approval-gate and draftcat for agents that send or write for real; graphiti-local for memory that knows what was true when.
| If you need to… | See how I approach it |
|---|---|
| Find the current customer decision across disconnected systems | Operational context layer — sources, timestamps and permissions travel with each answer |
| Let an agent act without giving it unchecked write access | draftcat and agent-approval-gate — proposals, human verdicts and audited dispatch |
| Keep track of what was true when | graphiti-local — a local temporal knowledge graph with read-only MCP retrieval and human-gated fact ingestion |
My merged contributions include RAG query validation before logging in Tether's QVAC and GPT-5/o-series vision support in WeKnora. See all merged PRs.
I also build creator tooling: capcut-cli is an independent CLI for editing CapCut/JianYing project files.
I publish implementation notes for people building agents beyond demos. Recent examples:
- HTTP If-Match and 412: stop stale agent writes after approval — atomic target-version checks preserve a newer human edit; includes a runnable HTTP race fixture.
- Permission-aware retrieval without count, title or provenance leaks
- Why an agent proposes facts but never writes directly to the knowledge graph
Series index — all 19 entries · Tested automation examples
I work with DACH teams that need AI to answer from their actual operational state and act within clear permissions. A first engagement can review one workflow's sources, identity, write path and failure modes before a phased build.
Discuss your workflow · Email me · Malt profile and references