We build AI products and the tooling that makes them cheap to run. Plain code where the work is predictable, local models where they are enough, a big model only when the job needs judgement.
The boring version is cheaper and breaks in ways you can predict.
We treat agent workflows like payment systems: bugs cost money, and so do tokens.
Work
-
The harness we run our own coding agents through
You write the goal. It plans, edits, tests and commits, using cheap models for the routine steps.
Coding agents forget everything between sessions, never show what a task cost, and check their own work. clauaid is our fix for all three. It keeps memory across sessions, sends routine edits to cheaper models, and lets tests decide whether an edit ships. The tests come from a separate role than the code, and a lint at commit enforces that. Big risky changes stop at a gate and need a written plan and an approval before any code gets written. We use it daily, on itself included.
-
A plugin that makes a coding agent read and write less
Sits inside a coding agent and stops it reading or writing more than the job needs.
An agent burns most of its tokens on reading, not on the answer. lean-harness is a small plugin, plain bash hooks, no dependencies. It pushes the agent to read less and build less: no whole-file reads when a grep would do, no extra code nobody asked for. And when something fails and later gets fixed, it writes the lesson down so the next session does not repeat it.
-
Scores the fit first, writes the application after
Reads a job posting, scores the fit against your record, and only then writes the application.
Mass applying is a waste, and hand-tailoring every resume eats days. You also end up pasting your whole work life into someone's model to do it. soshitsu scrubs personal data out before anything reaches an LLM and puts it back after. It scores each posting against one master profile, and every point in the score needs a quote from the record behind it. Under seventy percent fit, no application. Above it, a tailored single-column resume that parses cleanly in ATS systems, with nothing made up. FastAPI and Redis underneath.
-
Fine-tuning a small model without becoming an ML engineer
A platform where a creative, not an engineer, fine-tunes a small model on their own material.
A small model tuned on your own material keeps your voice better than prompting does, and costs less to run. But fine-tuning assumed an ML engineer at the keyboard, and creatives, who have the most distinctive material, usually do not have one. quickft removed that step. You upload your material in the browser, it runs a LoRA fine-tune through Predibase, and when it is done you chat with your own model. We got it to a working MVP and a Y Combinator application, then circumstances changed and we set it aside. The fine-tuning experience stayed with us.
-
A scouting pipeline for seed-stage Indian consumer brands
Watches thirty sources for young companies doing well and ranks the ones worth an early call.
By the time a young brand shows up in the funding databases the round is already crowded, and those databases cost real money anyway. The earlier signs are public but scattered across marketplaces, registries and Instagram. analyst watches about thirty of these sources, merges what it finds into one profile per company, drops anything already funded or too established, and scores the rest on signals like domain age and how coherent the brand looks. What comes out is a short ranked list an investor can actually call through. We ran it against a universe of around four hundred real brands.
-
The photo archive drafts its own posts
Photos and notes go in. It works out what belongs together and publishes the result.
The camera roll grows faster than anyone actually sorts it, and every publishing tool still wants you to sit down and operate it. kiroku flips the roles. You feed it photos and notes, it runs the pipeline on its own — dedup, aesthetic scoring, OCR, clustering — and drafts posts in your voice. Your part is just the decision: accept, park, or reject. Accepted posts go out on its built-in microblog.
-
Making an open coding agent run on a cheaper model
We forked a terminal coding agent to run Gemini and cut what it cost to run.
aider on Gemini was broken in very specific ways. The model would reply "please go ahead and apply the edits" instead of actually editing, failed retries looped forever, and sessions burned around three times the tokens they should have. We forked it and kept every fix in one isolated package: one call per phase that carries file operations and edits together, loop detection, and a patcher that tries exact match, then diff, then fuzzy before asking the model again. Token use came down to roughly a third of what it was.
-
One place to finish a reel
Scores, captions, grades and cleans a video in one place, instead of four separate tools.
Finishing one reel used to mean four tools: something for music, something for captions, a colour grade, an audio cleaner. reel studio does it in one web editor — procedural music, captions from a vision model, grading, dialogue cleanup. Everything runs locally on a 16GB laptop, so models load one at a time and get freed before the next. Nothing goes to a cloud. We built it for our own footage first.
Get in touch
If this sounds like the way you want your systems built, write to us.