AMR Screening
AI-assisted screening of approved medicines against four drug-resistant bacteria.
Software that holds its shape.
SNOWBROS is an independent engineering studio. We build developer tools, products and interfaces that tell you what they actually did.
Same input in. Same result out.
That is the rule we build by. A media player that names the decoder it used. An analyser that shows the evidence for every finding. A transfer that never touches a cloud. Software you can trust because it never pretends.
We are a small studio led by Ved Chauhan, working in Rust, TypeScript, Python and Flutter — from the engine to the last frame of the interface.
Selected work
Products we designed and engineered for clients, then tools of our own. All shipped in 2026.
-
-
NOVA
A privacy-first desktop media player: no account, no telemetry, works offline.
-
Accounic
A multi-currency ledger: web, Android and Windows clients on one Postgres schema.
-
SwiftDrop
Photos and files straight between phone and PC over your own Wi-Fi.
-
STC Academy
A teaching platform for CBSE and GSEB students, with an admin command centre behind it.
-
Atlas
Static analysis for JavaScript and TypeScript that reports only what it can prove.
Also shipped
- Creative EngineeringOur open-source agent skill and eight @snowbros libraries — see them run in the labTypeScript · three.js
- Lunora StudioHandmade-bouquet store with a scroll-told landing pageNext.js · Supabase · GSAP
- MentorAn engineering-mentor skill for Claude CodeAgent skill
What we build
Developer toolsAnalysers, command lines, language servers, editor extensions.
ProductsDesktop, mobile and web apps that one team can run.
Tauri and Flutter on the desktop and phone, Next.js and Supabase on the web, one schema behind every client. Release pipelines included, because shipping is part of the product.
Seen in NOVA, Accounic, SwiftDrop and STC Academy.
Interfaces and motionProduct UI, research dashboards, launch sites, interactive 3D.
Motion that explains what changed, interfaces that stay calm where work happens, and the occasional moment that people remember — like the name at the top of this page.
Seen in AMR Screening and on this page.
Bring us something that has to hold.
AMR Screening
AI-assisted screening of approved medicines against four drug-resistant bacteria.
The problem
Repurposing an approved drug is a shorter road to the clinic than discovering a new one. This prototype screens the FDA-approved library against MRSA, E. coli, K. pneumoniae and M. tuberculosis.
What we built
- A pipeline from ChEMBL bioactivity and FDA approvals to 1024-bit Morgan fingerprints and a benchmarked model zoo per pathogen.
- Docking of prioritised candidates with AutoDock Vina against one validated protein target per pathogen.
- A dashboard where every prediction can be inspected down to its evidence.
Decisions that mattered
- A stated scientific boundary on every screen: these are computational hypotheses, not evidence of efficacy.
- Numbers are counted once and explained, including the structures that were dropped and why.
1,761approved medicines screened
NOVA
A privacy-first desktop media player: no account, no telemetry, works offline.
The problem
A media player that tells you the truth: what it decoded, with which decoder, and where you stopped — across restarts and crashes, even when files move.
What we built
- Native video through libmpv with hardware decoding and software fallback, under a transparent React interface.
- Frame-accurate seeking, A–B loops, a library with background scanning and full-text search.
- Engine-confirmed state only: the interface shows what libmpv reports, never what it hopes happened.
Decisions that mattered
- Pure domain rules in their own Rust crates, so the queue and resume policy are tested without an engine.
- Professional formats verified on synthetic samples — ProRes, DNxHR, HDR10, 8K — before anyone trusts it.
1 frameseek precision on the timeline
Accounic
A multi-currency ledger: web, Android and Windows clients on one Postgres schema.
The problem
It answers four questions and little else: who do I have an account with, how much do they owe me, how much do I owe them, and what has been settled.
What we built
- One backend where the database owns every balance: no client ever adds up a column.
- A FIFO settlement allocator, and entries kept in the currency they were entered in.
- A Windows installer and an Android APK that check GitHub Releases and offer the new build.
Decisions that mattered
- Forced row-level security on every table, verified by 71 tenant-isolation checks over real HTTP.
- Releases are built by CI and left as drafts: a person decides when a build reaches installed clients.
234SQL assertions in CI
SwiftDrop
Photos and files straight between phone and PC over your own Wi-Fi.
The problem
No cloud, no cable, no account and nothing to install on an iPhone: the phone scans a code, the PC allows it, and files travel in parallel chunks over the local network.
What we built
- Native Windows and Android apps in Flutter, plus a browser path for iPhone.
- Phone-to-phone transfer over WebRTC, where the code carries connection details and never data.
- PC-to-phone downloads served straight from disk, with nothing copied.
Decisions that mattered
- Works without internet: only two devices and a router or hotspot.
- Firewall trouble is the commonest failure, so the fix ships with the app.
STC Academy
A teaching platform for CBSE and GSEB students, with an admin command centre behind it.
The problem
Students from primary school to HSC, on two boards, with tracks for JEE and NEET — and a small team that has to run admissions, attendance, fees and materials without spreadsheets.
What we built
- A dual-pathway course system: board-aligned curriculum alongside competitive-exam preparation.
- Faculty mentorship profiles and small-batch enrolment, filterable by level, subject and pathway.
- An admin command centre for students, teachers, subjects, attendance, QR codes and fee status.
Decisions that mattered
- Supabase for auth, storage and real-time data, so the school runs on one managed backend.
- The same records serve the public site and the admin tools: nothing is entered twice.
Atlas
Static analysis for JavaScript and TypeScript that reports only what it can prove.
The problem
Linters guess. Atlas maps a whole project — every import, export, env var and framework boundary — and reports circular imports, dead files, server code leaking into client components and hardcoded secrets, each with the chain of evidence that produced it.
What we built
- A native Rust engine with a shared semantic IR and a tree-sitter frontend per language.
- A command line (
npx snowbros analyze), a built-in language server and a VS Code extension. - Deterministic output: the same codebase in gives the same findings out, every run.
Decisions that mattered
- No model decides whether an issue exists. Every finding is provable or it is not reported.
- Speed is a feature: on axios (431 files) a cold run takes about 230 ms, a warm one about 76 ms.
230 mscold run on axios, 431 files