Projects

The two products at the top are still private. Everything after them links to code, a specification, or documentation that can be inspected directly.

In development

Expense splitting records every change as an event instead of replacing the current bill. Concurrent edits both survive, balances can be rebuilt from history, and property-based tests compare the settlement algorithm with a simpler reference implementation.

Receipt scanning turns a photo into line items that several people can claim at the same time. It runs entirely on Cloudflare Workers. The extraction pipeline separates image reading from later stages, can switch model vendors at each boundary, and rejects results whose arithmetic does not match the receipt total.

Developer platforms and TON

tonapi-js is the generated TypeScript client for TonAPI. I maintained it alone from mid-2024 through March 2026, rewrote its packages, and built the generator that maps OpenAPI int64 values to bigint and produces the same client for browsers and Node with one runtime dependency.

ton-console is Tonkeeper’s developer console. I was its only developer for 21 months. The work includes billing moved from TON to USDT without a hard cutover, webhooks management, pricing tiers, analytics queries and airdrop tooling.

TON Connect is the protocol used by apps and TON wallets to connect. Two ambiguities I found became specification changes: the address format in a sign-data response and the encoding of appDomain.

ton-sign-data-reference is a compact TypeScript implementation of TON Connect SignData. It covers text, binary and TON-cell payloads and is intended to be read alongside the specification.

opentonapi contains TonAPI’s public OpenAPI schema and Go backend. My contribution was almost entirely to the schema: 64-bit integers in JavaScript, date formats and response types. I made two small fixes in the Go code.

tonconsole-docs contains integration guides I wrote on transaction tracking, data signing and hash normalization.

This site

mois.pro builds this page, two CV pages and two PDFs from one repository. Pandoc renders the Markdown, Typst produces the PDFs, and a build check reads their text layer to make sure dates still belong to the correct jobs.