← Back · Everything, in depth
Two kinds of work. Open source you can read, install and judge for yourself. Client and platform work you cannot, described by what it does and what it had to get right.
Every answer the company already has, in one place that never makes one up. Ask how the work is actually going and it replies in plain language with the receipts attached, to people and to AI agents alike.
Shows you only what you already have the right to see, and says I don't know rather than filling the gap.
Automation the whole company can build. Assemble the AI and integration steps you need, press run, and watch it work. What used to be a ticket to engineering became an afternoon.
A catalogue of enterprise software, made AI-native without rewriting any of it. One small layer slots in behind products built years ago and gives every one of them a modern assistant.
Weeks per product became days, and nothing the customer already relied on had to change.
Hundreds of pages of financial records per case, read and reconciled in minutes instead of days. The kind of review that used to eat a paralegal's week.
Live in production with paying customers.
Give it the brief, get the cut. Point it at a film, documentary or podcast, say what you are looking for in plain English, and it comes back with the moments worth posting.
Judged better than the leading commercial tool in a head-to-head on a feature documentary.
A first-pass reviewer for pharmaceutical quality packages that never waves anything through. It reads the whole evidence pack, catches what does not line up, and shows its working.
105/105 and 326/326 on both ground-truth benchmarks, delivered in under a day.
Sales practice against an AI that pushes back. Reps call in, run the conversation for real, and get scored on the four things that decide the deal.
It can clone the exact customer who beat them last week and let them try again.
Feedback people actually give. Instead of a form nobody fills in, everyone at the event gets their own line to call and simply talk about how a colleague did.
Ran live across a four-day conference for every attendee.
Ask the company archive a question and get an answer, in whichever language you asked in. Search, summarise and translate across everything the business has ever filed.
Shortlists that understand the role, not just the keywords. Write the job description, drop in the CVs, and get back the people who genuinely fit.
Your AI coding assistant, except it already knows your codebase. No more burning half a session re-reading your own files: ask where something lives or what calls it and the answer comes back instantly.
Runs entirely on your machine. No API keys, no network, no telemetry.
Give your agent fifty tools without paying for fifty tools. It fetches only the handful the job actually needs, and leaves the rest out of the conversation.
Reclaims the slice of the context window that used to disappear before the first message.
An AI security operator you can genuinely let loose. Point it at a system you are authorised to test and it plans and runs the engagement itself.
Hard limits it cannot talk its way around, and a full audit trail of everything it did.
Stop trusting eval scores that quietly drifted. It keeps your AI judge honest against real human labels, and tells you when a result is real rather than noise.
When the judge moves, it refuses to give you a number instead of handing you one you would have trusted.
Safety that runs on the device, not on the bill. Tiny models that catch jailbreaks and harmful content before they reach your LLM.
20+ languages, under 50ms, on ordinary hardware.
Login that proves itself. It wires up sign-in and then actually signs in to check, so you find out it works now rather than in production.
When it fails it names which side is wrong and what to change.
Smaller tools, same year
cat file | tok.TypeScript
promptsizesize-limit for prompts: set a context budget and fail CI when a prompt outgrows it.TypeScript
passmusterA verify-and-retry loop: generate, check against a schema or judge, retry with feedback.TypeScript
agentvuRecord an agent's run and replay it as a live terminal dashboard.TypeScript
heyllmTalk to an LLM from the terminal. git diff | ask 'write a commit message'.TypeScript
readmecastTurn a Markdown file into an animated terminal demo as SVG, with no tape files.TypeScript
envvetValidate a .env against an .env.example contract, locally and in CI.TypeScript
gitbroomInteractively sweep stale local branches: merged, or whose remote is gone.TypeScript
shiptagGenerate a shareable SVG card for any repository.TypeScript