What has my attention right now, and roughly why.
Integration and data-pipeline engineering: CRM workflow automation, ingestion services and scheduled jobs on managed cloud infrastructure, and warehouse-backed reporting behind access control. One of those services is LLM-assisted — it scores call transcripts against a rubric, deterministic checks first, human review after, with the model never given the last word. Categorical on purpose: the interesting part is the shape of the work, not whose it is.
A lossless compression proxy for MCP tool output — 59 percent fewer tokens, with the original forwarded untouched whenever a perfect round trip cannot be proven. Recent work has been on the security edges: credentials must not survive a cross-origin redirect, and a config generator that measures performance is never allowed to overwrite a safety setting. Case study.
Given a diff, it names the user journeys the change could have broken, walking a static call graph rather than a coverage report — so it answers on a repository that has never run its tests. Just went public and onto PyPI. The current work is the honest half: joining what it selected against what running the journey then found, so the miss rate becomes a number instead of an assertion. Case study.
An agent harness whose verdict on a change is four deterministic checks with no model anywhere in the judge, and no auto-commit at the end of them. The loop that drives it still runs in someone else’s runtime; moving it in — with a write-capable tool set behind an isolation contract — is what is being built. Case study.
A personal knowledge base on Postgres that agents write into by proposal only, with every proposal reviewed before it lands. This quarter it moved off local embeddings onto a sidecar — production image down from 2.14 GB to 442 MB — and shed half its tool surface, because response bytes turned out to be the real constraint, not tool count.
Spatial modeling over fifty thousand Michigan well records, and the contractor-intake business built around it. Depth is predictable from location; yield is not. Both of those are useful to know, and the second one took longer to accept.
Ten machines under Proxmox carrying real routing, monitoring, and backup — plus the recurring lesson that ordering and detection are not controls. A VPN-bound workload gets pinned to the tunnel interface by name, because restarting things in the right order only works until it doesn't.
Email is the best way to reach me — [email protected].
