A short note on what we built between June and August. A rebuild that paid for itself, a media product shipping every week, and a capability we can drop into your product.
We are an AI product studio. We design, build, and ship AI products and the engineering underneath them, in production, fast. Here is the last three months in one page.
The quarter in brief
Case study · Client engine rebuild
A client of ours generates AI music videos at volume. The engine that served them was written in Java and had grown brittle. Reverts were frequent, the mobile apps broke in ways nobody caught until users did, and the cloud bill climbed with every view.
We rebuilt the registry from Java to Python, contract-first, and verified it against the real mobile clients before a line shipped. The result was not just cleaner code. It was a service that costs a fraction to run and stopped breaking.
The egress win came from moving an immutable, write-once media set onto zero-egress storage. On the old stack that same traffic was billed at cloud egress rates, a cost that doubles as vendor lock-in. Both mobile apps verified green before launch, because we tested against the actual client decoders instead of trusting a green status code.
Capability · Generative video
For a client, we build a consumer app that turns a song into a video you star in. This quarter it grew from a demo into a product with a weekly rhythm. We treat still-image design and motion as separate decisions: lock identity and story on approved references first, then let the video model perform the shot. It keeps quality high while the pace stays fast.
The lesson we keep proving: approve the still before you pay for the motion. It is the difference between a product that ships weekly and one that stalls in revisions.
Capability · Live collaboration
Sometimes two people need to be on the same screen at the same time. A support agent guiding a customer. A team reviewing a design together. A seller walking a prospect through a flow. We have built the engineering for this and can add it to your product.
It runs peer to peer, so cursors, edits, and presence move in real time without a heavy media server. Small rooms need almost no infrastructure. When you outgrow that, the same design graduates to a forwarding server without a rewrite.
We ship it with the parts that usually get skipped: presence, reconnect handling, and concurrency that does not throw away someone's work when two people touch the same thing. If live collaboration is on your roadmap, this is weeks of work we have already done.
From our notebook
We publish the useful parts of how we work. A few pieces from this quarter, open to read:
Products · try what we build
We build for clients, and we run our own products alongside that work. Three you can use today, free to try:
Meeting notes that write themselves. Steno captures the conversation and hands you clean, searchable notes when it ends.
Visit Steno →Push HTML, docs, images, and PDFs straight from your coding agents to live, commentable, versioned pages your team actually reads. This newsletter runs on it.
Visit thing →A tiny floating cat that watches your Claude Code and Codex agents from a corner of your screen. Live context meters, permission approvals, one cat.
Visit Tablo →If you are shipping an AI product, weighing a rebuild, or want live collaboration in your app, let's talk. A short call, no pitch deck required.
Book a 30-minute call → Or just reply to this email. It reaches us directly. kapil@unravel.tech