-
A voice-first assistant for people with effortful speech (built for my dad, who has Parkinson's). Reminders, medication timing, calls for help. Real-audio evals: 58% baseline comprehension → 82% with repair. 656 tests.
-
My agent-orchestration stack, open-sourced: memory architecture, meta-learning loops, playbooks that let one operator direct many agents. v0.5.0, 8 production skills.
-
Metadata-only receipts and fail-closed tripwires for agent trajectories. No content capture, zero deps, CI-gated.
-
Shift marketplace and orchestration for the frontline workforce: matching, credentialing, scheduling, pay. Built with UKG Labs as partner and investor.
Founder, CTO, agent builder
I build small teams of AI agents that ship real software.
01Building
4 active
02How I build
- planPlan before code. The design doc comes first. If I can't write it down, it isn't ready to build.
- buildAgents build, I direct. Coding agents write the code. I set the spec, review the diff, and own the decision.
- reviewCross-model review by default. Every change is critiqued by a different model family before it ships.
- measureMeasured numbers only. Evals run with negative controls. I report what's measured, never what I'm hoping for.
- guardFailures become guardrails. Each failure turns into an executable check, not an apology.
- queueA night queue. Agents build while I sleep. Morning is for review.
03Build log
- Jul 2026Shipped clawrari.com: brand, landing, and story pages, three review passes across two model familiesclawrari.com ↗
- Jul 2026Parker real-audio eval on dysarthria corpora: 58% → 82% comprehension with repair loopnote →
- Jul 2026trace-to-tripwire v0.1: fail-closed CI gate for agent trajectory receiptsgithub ↗
- Jul 2026Two agents (Claw + Hermes) now build and review each other's work over a private channelnote →
- Jun 2026Cross-agent eval framework: strict scorecards with negative controlsgithub ↗
04Notes
10 entries
- Jul 2026 my own repo disproved my readme I caught my own README overclaiming Parker, and the real eval number was better than the lie.
- Jul 2026 a repo can commit daily and still rot My agent audited our open-source sync and found three months of drift behind green CI and a daily commit streak.
- Jul 2026 guardrails, not prompts A prompt tweak fixes one bad output; a guardrail in the loop stops every one that fails the same way.
- Jul 2026 managing coding agents is managing juniors who forget everything overnight Coding agents are capable juniors with no overnight memory, so the work is in the brief, not the model.
- Jul 2026 a subagent dies quietly: the context-death contract Five subagents died in two weeks, all context blowouts, so the anti-death rules became a required contract on every spawn.
- Jul 2026 metadata-only receipts: trace, then tripwire Every agent run leaves a metadata-only trace that a fail-closed tripwire gates on before the next run proceeds.
- Jun 2026 the night build queue I set the plan at 10:30pm and agents build while I sleep; it only works if every run re-derives its numbers by morning.
- Jun 2026 two agents reviewing each other over a private channel Two agents on two machines reviewing each other produced our most novel work, and a six-hour deadlock that taught the most.
- May 2026 cross-model review is the default Every change gets reviewed by a different model family before it ships: one writes, one reviews, single round.
- May 2026 memory is the product surface For an agent, memory isn't a bolt-on feature; it gets a measured architecture with benchmarked recall and negative controls.