Work
Each card: the problem, what we built, one proof, and where to see it. More are added as each one is checked.
Data migration · preview
Problem. Postgres→ClickHouse copies drift without warning: deleted rows stay and long text arrives empty.
What we built. A CLI that installs ClickHouse, Keeper and PeerDB over SSH, then copies the data and checks it by checksum before it allows cutover.
Proof. 15 Postgres tables (5 hot), 505,800 rows, snapshot to ClickHouse in 23 s. All 15 tables equal by checksum (64 key ranges per table, 0 differ).
Agent security · open source
Problem. API keys you paste, or that tools print, go straight to the model provider.
What we built. A Claude Code mod and opencode plugin. It swaps each secret for a $NAME on your machine and fills the value back in only when a command runs.
Proof. 232 patterns (217 of them from the gitleaks ruleset) turn each secret into a stable name.
Agent guardrails
Problem. Coding agents run destructive commands and obey injected instructions that keyword rules miss.
What we built. Small local yes/no checks inside Claude Code, Codex and your own agents. They run on CPU next to rules, never alone.
Proof. 42 of 60 injected instructions caught by the local model, against 20 of 60 for a keyword rule; 1 false alarm in 56 clean inputs. The same person wrote parts of the test sets and the rules.
Postgres backups · MIT
Problem. App teams without a DBA find out the backups never worked on the day they need one.
What we built. A Postgres extension and CLI. It backs up every database daily by default to your own S3 bucket, and new databases within a minute; one command restores.
Proof. Tested on Postgres 13–18. MIT licensed, with a free AWS Marketplace image.
Model serving on AWS
Problem. Teams want to serve their own decision model in their own AWS account, without per-call fees.
What we built. A container on ECS Fargate that serves an OpenJevX-format model from your S3 bucket, with hot reload and an API key on every call.
Proof. Small image (about 106 MB), CPU only, no per-call fees.
Hardware · early access
Problem. Putting phone calls in a product means per-minute cloud telephony and a carrier in the middle.
What we built. A box with your own SIM. An AI agent or a person makes and answers calls and texts through it, with open software. The board design is drawn; units are not built yet.
Proof. A 2-minute, 3-turn AI call on a live mobile network.
Developer education
Problem. Developer tools are hard to grasp from a README alone.
What we built. Short narrated demos and explainers of our tools and of agent-engineering topics.
Proof. 21 videos on the channel.
See consulting for what we take on and prices in 10 seconds.
