Bitnet.ge
Built by me, live today: the new identity checks, card payments, bank-app deposits, and a rule that each customer action runs only once, even after a double tap.
A crypto exchange licensed by the National Bank of Georgia.
Saba Janelidze Principal Engineer · Fintech & AI
AI agents for teams that move money, without losing a payment.
I help fintech teams use AI coding agents safely, and I build the AI features their product needs.
Since 2023 I've helped build a crypto exchange licensed by the National Bank of Georgia. I lead its new core platform's architecture and wrote most of the team's AI coding set-up.
AI agents can write this, and it can pass tests. My review agents check money code.
- Balance
- 20
- Paid out
- 160
- Exchange short
- 80
Each run is a recorded trace of the game's real code, checked by a test on every build.
- 220,012 registered users on the exchange I help build exchange's registered users bitnet.ge, Oct 2026
- 2,574,508 transactions on the same exchange exchange's transactions bitnet.ge, Oct 2026
- 237 of 278 architecture decision records for its new core platform, first-authored by me architecture decisions I first-authored platform code, Oct 2026
- 38 of 38 customer actions that must declare how a repeat is handled, or the build fails customer actions with a build-checked repeat rule my rule, since Sep 2026
Selected work
Playtime.ge
Live listings · booking built
19 padel clubs in Tbilisi listed publicly. Booking with payment is built; payments are still in test mode.
WasteAway
Pre-launch · website live
A live website and a worked-out cost model. Next: the first pilot building.
Chama
Past work
Designed and wrote the backend for live dynamic pricing.
UniHub
Live · on hold
Schedule, exams, grades and shared lecture notes in one place.
What I do
Three ways I can help teams that move money.
-
AI coding agents in your repo, set up safely
First an audit, then a 2-week rollout: coding agents in your repo with rules, safety hooks and automatic checks, and training for your team.
In one audit on my team, a verifier agent re-checked 111 claims made by other agents and corrected 12. I wrote most of that set-up: 6 review agents, 13 safety hooks, and a build check on the set-up itself. How I work
-
AI features in your product
An assistant, document search or back-office automation, with tests for its answers, logs, cost limits, and a person's OK before it changes data.
The same rules as my own work: separate review agents, and a person's OK before anything touches money or data. Smallest proof today: DevNews, my open-source AI reader (three providers, with a fallback when one fails). Code on GitHub
-
Fintech & payment systems
Payments, identity checks, audit history and staff tools, built so a request that arrives twice runs only once.
Live on the exchange I help build: my payment-provider integration, with bank-app deposits, card payments and recurring card charges. The work
How I work with AI agents
Agents write most of the code. Automatic checks block what fails. A person decides on money and anything public.
- 1
Plan and rules
Every task starts from a written plan and the project's rules, which the agents read first.
147 entries in a behaviour contract that the build checks PlayTime, Oct 2026
- 2
Agents write the code
Each agent works on one task, on its own branch.
7 features shipped live in 4 days PlayTime, Oct 2026
- 3
Automatic checks
Lint, types, tests and schema checks block anything that fails.
27 automatic checks in 12 pipeline files PlayTime, Oct 2026
- 4
Separate review
Separate review agents check the risky parts: money, permissions, data.
111 agent claims re-checked in one audit; 12 corrected my fintech team, 2026
- 5
A person approves
Money, accounts and anything public need a person's yes first.
All 11 launch decisions made by me, not an agent WasteAway, Oct 2026
About
I'm a Principal Engineer based in Georgia. Since January 2023 I've worked on a licensed crypto exchange. I started on its customer web app, moved to its backend, and now lead the architecture of its new core platform.
On the side I build my own products, PlayTime and WasteAway, with AI agents doing most of the coding.
Open source
- Arcana: 18 skills and 5 agents for Claude Code and Codex. 2,420 npm downloads in the last year.
- acctswap: a Rust tool that switches Claude accounts across sessions. A TLA+ model, Kani proofs, and 435 of 435 planted bugs caught by its tests.
- OrgClone and UserClone: clone a whole GitHub org, or all your repos onto a new computer. 1,982 npm downloads in the last year.
npm counts, Oct 2026: 4,402 across my three packages, bots and CI servers included.
Stack
- Languages
- TypeScript, Go, SQL, Rust (tools)
- Backend
- NestJS, GraphQL, PostgreSQL, Redis, Job queues, OpenTelemetry
- Frontend & mobile
- React, Next.js, Astro, Tailwind, Expo
- Cloud
- Cloudflare Workers & D1, Vercel, Railway, GitHub Actions
- AI
- Claude Code, Codex, LLM APIs with provider fallback
Let's talk
Tell me what your team is building, how you use AI agents today, and where you don't trust them yet. A short intro call is the easiest start: email me and we'll find a time.