· 3 min · updated Sep 27
building
titles travel badly. these are the things themselves — what each one is, what i built, and where it stands.
In this note03 sections
evidane
evidence-backed legal ai. eighty-seven specialist agents read the document and argue over what they found, a ten-pass pipeline verifies every finding, and a lawyer signs off before anything leaves. a finding you cannot trace back to its clause is a finding you cannot use, so evidane never shows one.
- live at evidane.com (opens in a new tab), with 1,000+ signups and paying customers
- five us law firms working with it
- v1.0, then 1,000+ commits in fifteen weeks
brekfuz
the ai agent platform over a company's knowledge, backed by pareto holdings, pear vc, z fellows, cory levy and collective global. i built every service, the queues, the infrastructure. answers land in the thread where the question was asked, split by who is cleared to read each source.
the memory agents are the heart of it. they extract, consolidate and recall what the company knows, and they have served 100,000+ requests. underneath, every source feeds its own ingest stream with fair per-tenant admission: with ten busy tenants, a new one waits behind at most 576 jobs, never behind someone else's 100,000-job backfill.
then breken (opens in a new tab), brekfuz's second product: self-healing software. it maps a codebase, proves a bug is real, and opens the pull request that fixes it. i built every layer of it — the code intelligence, the renderer and the sandbox — and maintainers at langflow, tailscale, unsloth and twenty have merged its fixes.
- four production services on aws: web, api, a worker fleet and a go memory service
- 100,000+ requests served by the memory agents
- 213,761 production requests in fourteen days, counted at the load balancer
- row-level security forced in postgres and verified at boot
- breken: 54 fixes merged in 34 repositories, listed at breken.ai/open-source (opens in a new tab)
the open-source shelf
five projects i built in the open, each one a question i could not stop asking.
- helmryth (opens in a new tab) — a private operating system for autonomous work: persistent operators on the clis you already own, and nothing consequential runs until you answer
- run phantom (opens in a new tab) — see the run, find the reason: a local-first debugger that turns agent traces into timelines, comparisons and regression checks for ci
- jitmind (opens in a new tab) — memory for long-running agents that knows when a fact stopped being true
- fablewright (opens in a new tab) — risk-gated orchestration for claude code and codex, where nobody grades their own homework
- orrery (opens in a new tab) — architect-first orchestration on a zero-dependency mcp server, every model pin checked against what actually ran
and in other people's code: 350+ pull requests to 50+ projects, including hugging face, superset, burn, baseten and mcp-use. counting breken's, 100+ fixes are merged upstream, into projects with 800k+ combined github stars.
the reasoning underneath these is in how i think; what i am on right now is in now.