Long-reads on how CricCuts cuts your highlights automatically, why on-device AI is the future, and the story behind the app — written by the person who builds it, for a game where good players get seen, coaches reach further, and sport finds its way into more lives. Our mission & vision →
Why not let the phone translate our cricket physics course into whatever language a student speaks? We tried it on all 1,102 lines of the Science Lab, in Hindi and Tamil. It took 23 seconds, offline. It also switched off the pitch, turned the batter into thunder and flipped the answer to a quiz question. What went wrong, why our own careful translations had the opposite problem, why we're writing the lessons in Tanglish and Hinglish instead, a checklist if you're trying the same, and the open data.
LaunchCricCuts is in open beta. A boy heading home from nets with one bar of signal and no data left should still be able to cut his own highlights before he gets off — and that sentence, not a privacy policy, is what decided the architecture. What the app does, what the twelve animated scenes are for (coaches and social are different jobs), why the effects layer is still experimental, and the engineering opinions underneath: a renderer a machine can check, constants calibrated against human-curated footage instead of chosen by eye, offline harnesses that render art with no device attached, and a scene pipeline that hands each step to a different model. Plus the scene that passed every test and looked like a mole.
GuideGold, blue, red or white: the lines that gather around the batter's head before contact are coloured by how still that head stayed from stance to strike. The colour key, why "keep your head still" follows every batter from their first net, what the colours can't tell you, how to re-check a reading that looks wrong, and how to put the key on your exported video.
ArchitectureA renderer that returns a picture can be interrogated by a machine; one that draws into a system cannot. That single distinction decided how eleven of our twelve animated scenes are built — and it is the same rule behind ports and adapters, functional core / imperative shell, SQLite's storage interface, FoundationDB writing a deterministic simulator before writing the database, and React returning a description of the UI instead of mutating the page. The full technique is here (ARGB integer arrays, alpha compositing, the mobile animation budget) along with an honest comparison against Canvas, GPU shaders, Lottie and Skiko — but the argument is that the principle is thirty years old and only its price has changed.
Product & UXMost "AI video editors" hand you a prompt box: describe what you want, wait, and re-prompt if it's wrong. CricCuts is built around a different unit of collaboration — the anchor, a single AI measurement of when and where something happened, shared by dozens of effects, that you correct with one pin or one drawn box instead of a hundred small edits. What that looks like concretely, why cricket forces it, and an honest look — after actually checking — at whether anyone else already builds this way.
How it's builtHow CricCuts was built. It started as a bake-off between Google Antigravity, Android Studio's Gemini agent and Claude — and the forensic trail survived, right down to the hour Antigravity spent scaffolding the app and the twelve minutes it took to hand over. 267 commits, a commit-level answer to "which model actually wrote this", and the cricket facts that overruled every one of them: 25 of 30 bat-ball contacts make no sound, your neighbour's crack is louder than yours, and the fullest swing in cricket is a miss.
India @ 80The British split Indian cricket into teams by religion — Europeans, Hindus, Muslims, Parsis, "The Rest." Eighty years into independence, the game turned into one of the country's biggest unifiers instead, and did the same for the West Indies and South Africa. The history, the numbers behind cricket's rise from a 1982 colour telecast to a global audience rivalling the Super Bowl, and why we built CricCuts as India's next gift back to the game.
How it's builtWe added a frivolous thing — a comic-book gust of wind whooshing off the bat at contact. Making it land on the right frame, pointing the right way, exposed a timing error every accuracy score we had was structurally blind to: the loudest sound near a cricket shot is very often not the shot. Fixing it gave us a better way to track the batter and a measurably more accurate engine. Plus six hundred years of art dragging technology forward — perspective, the Jacquard loom, Muybridge's horse, and the cartoon that paid for modern AI.
The scienceA bat meeting a ball is over in a thousandth of a second — so how does a phone catch it? A plain-English field guide to the ideas that make it possible: the physics of the sound (timbre, onset, attack, the envelope, the spectrum) and the geometry of the motion (luma grids, motion energy, the repetition heatmap, and the stance→trigger→swing→follow-through signature of a stroke). Concepts, not code — the vocabulary behind how CricCuts hears and sees the game.
Product & UXA general trimmer makes you scrub an hour of footage to find six balls. Cricket has a shape a purpose-built editor can lean on — so every thumbnail is framed on the moment of contact, the app reads adjacent nets and a batter who taps and wanders the crease, and a loading screen actually teaches you the game. The UX and product decisions behind CricCuts — plus TokTok vs coaching cuts, watermarked exports that get you noticed, why we're deleting our own advanced filters, and the one boundary box we'll never remove.
How it's builtWe build CricCuts with an AI agent in the loop — yet our single biggest accuracy breakthrough, learning where a recording's action keeps repeating, came from a human hunch no language model suggested. Not after one review, and not after many, by more than one of them — Claude and Fable included. On the real division of labour between a person and an AI. With a heatmap you can read at a glance.
The scienceYou can't fully feel your own technique — but a camera can show it to you, and your brain is built, right down to the neuron, to learn from what it sees. The science of self-observation — mirror neurons, prediction error, video self-modeling — and the exact cricket mistakes video fixes in batting and bowling, with real examples from Tendulkar, Kohli, Smith and remodelled bowling actions.
How it's builtThe cleverest thing about CricCuts is something you'll never run. We use LLM-powered agentic development to build, test and endlessly refine a complex, deterministic engine — the kind of app that used to need a whole team — and ship something cheap, instant, private and incapable of hallucinating. Plus a visual map of the loop we iterate in.
Interactive courseA guided, interactive course on the signal processing and small on-device AI models — MediaPipe Pose, Vosk and Silero VAD — that turn a raw net or match recording into a clean reel, automatically and entirely on your phone. Twelve modules, live demos you can poke, and a quick quiz on each.
Big pictureThe industry spent a decade making models bigger and more remote. The next decade belongs to small, specialised models running right on your device — faster, more private, greener and free. Here's why, and how CricCuts proves it.
Big pictureNot one giant brain, but a society of small specialists cooperating — the architecture of the human nervous system, and an emerging pattern in modern AI: mixture-of-experts, agent orchestration, and tools like Claude Code that run agents locally while leaning on the cloud. Graduate-level, in plain English.
Mission & makerCricCuts is an independent, one-person project — built in late-night coding spells between early-morning matches, fuelled by billions of AI tokens and a few hundred cups of black filter coffee. The story behind the app, the four commitments it exists for, and the sports media platform it is being built towards.
