Most of my work sits at the same seam: a model that is confident, and a system that has to be correct. The interesting engineering is almost never the prompt. It is the budget cap, the approval gate, the audit log, the fallback path, and knowing which parts of the problem the model should not be touching at all.
A few things I keep coming back to:
- Let the model translate, not compute. In
calpt, an LLM turns English into a command and SymPy does every piece of arithmetic — because a language model that is 97% right about calculus is a language model that is wrong in a way you will not notice. - Autonomy needs brakes before it needs features.
trentruns multi-agent workflows behind hard spend caps, an approval queue, a kill switch, and a tamper-evident hash chain over every state change. - Say what is actually proven. trent's README opens with a status table that says Not-yet-wired in three places, next to the exact command that would prove otherwise.
| Project | What it is | Why it's interesting |
|---|---|---|
| KoSh (case study) | Predictive dispatch for trucking fleets — risk scoring, HOS-legal recovery options, a learning loop behind safety gates | A seeded evaluation lab, a 9-metric scorecard, and shadow → canary → auto-rollback promotion. The eval evidence is itself treated as untrusted: stale, dirty-tree or commit-mismatched artifacts are refused |
| trent | AI cofounder OS — planner builds a DAG, specialist agents execute, a critic loops, a consolidator writes the brief | Approval-first architecture, per-agent token budgets, SHA-256 audit chain. 1,731 tests, 0 typecheck errors |
| ai-pipeline-builder | A Claude Code Skill for 30+ step autonomous pipelines | Steps are data, not code paths — so a run that dies at 19 of 30 says so, and the 18 outputs before it survive |
| calpt | A stock TI-84 doing real calculus over a phone hotspot | SymPy computes, the LLM only translates. Character-whitelist input validation, pinned TLS on an ESP32-C3 |
| casems-v1 | AI-augmented law-enforcement case management | Voice-to-report and court-readiness scoring with NIBRS-compatible structured output |
| stash | Second brain for ephemeral content | Vision-capable enrichment that degrades to a heuristic fallback instead of failing |
| Trench-OS | Content OS for creators | A brand-voice model extracted from real samples conditions every generation |
| Project | What it is |
|---|---|
| RudeStorm | Passive RF sensing — turns the WiFi already in a room into a sensor. Deploys no new primary sensors and transmits nothing |
| rf433-gateway | Standalone ESP32 433 MHz decoder and replayer — 56 protocol families, hundreds of real devices, no cloud and no home WiFi required |
Ten small, sharp defensive tools, each rebuilt to one standard — shared rich
CLI, hardened network surfaces, safe-by-default flags, 275 tests passing
across the set. Built to make the right thing easy and the wrong thing hard.
byte-bouncer · phish-pond · lantern-lan · threat-tea · key-kettle · caesar-salad · floodlight · lockbox-lane · sentry-sundae · patch-peeker
Every web UI binds to loopback. Every active scan is gated behind an explicit authorization flag. No exploits, no payloads.