September 6, 2026 · 3 min · updated Sep 27

about me

hi, i'm divyam. i founded evidane (opens in a new tab), i'm the founding engineer at brekfuz, and i ship: 1,600+ merged pull requests across my work and open source.

i take an idea all the way through: the architecture, the backend, the agents, the infrastructure and the launch. at brekfuz that was all of it; at evidane, everything a paying customer touches. evidane has 1,000+ signups, paying customers and five us law firms working with it. brekfuz is backed by pareto holdings, pear vc, z fellows, cory levy and collective global.

build the whole thing

the parts i care about are the ones a demo never shows. a queue that stays fair when one customer floods it. memory that survives a conversation and respects who is asking. a fix that proves itself before anyone reviews it. that is what decides whether a product works on an ordinary tuesday.

at brekfuz i built the ai agent platform — every service, the queues, the infrastructure, and the memory agents that have served 100,000+ requests — then breken, which finds a bug, proves it and opens the pull request that fixes it. at evidane, every legal finding traces back to the clause that earned it.

leave a trail

an answer should come with something you can check.

that is the thread through all of it: provenance in retrieval, citations in legal review, a proof attached to every fix, and a record of what an agent was allowed to do. anyone using the system can follow its reasoning back to the source, especially when the answer looks suspiciously confident.

the work here is written the same way: the decisions, the tradeoffs and the numbers, each tied to the system it came from.

learn with the code open

i learn by building the thing i am trying to understand. helmryth (opens in a new tab) runs persistent ai operators on your own machine, with a gate in front of every consequential action. run phantom (opens in a new tab) is a local debugger for agent runs. jitmind (opens in a new tab) is memory that knows when a fact stopped being true. fablewright (opens in a new tab) and orrery (opens in a new tab) keep a human architect in charge of coding agents, with pinned model lanes and a fresh reviewer that answers ship, fix-first or rethink.

i also spend a lot of time in other people's code: 350+ pull requests, and 100+ fixes merged upstream, mine and breken's. one favourite is a fix merged into burn (opens in a new tab) that stopped average pooling from counting padded zeros at tensor edges. small defect, wrong numbers everywhere downstream.

where to go next

for the roles and dates, read experience. for the projects, start with building. the ideas i have put into words are in writing, and this month's focus is in now.

if you want to know how i decide any of it, read how i think. if you would rather just talk, hit me up.

elsewhere: github (opens in a new tab) · x (opens in a new tab) · linkedin (opens in a new tab).