Software engineer in Istanbul. I build products end to end in Python and TypeScript, and for the last two years mostly products with LLM agents inside them: the billing, the evals and the deterministic checks that turn an agent demo into something people pay for.
Currently at Swarm Network, remote.
| What it is | My part | |
|---|---|---|
| Eclipse FM | AI-native music platform. People live-code music while AI agents compose alongside them. | Built end to end: React, Express, Python. Multi-agent composition system (a manager agent over four instrument specialists), credit and subscription billing, data model. |
| 43 | AI prediction platform. Agents reason over each question and publish their forecasts in public. | Built end to end: Next.js, TypeScript, Python. Web app, ingest bot, credit economy, backtesting harness, and the settlement pipeline with 13 code gates between the LLM judge and the database write. |
| Vault SDK | Secrets manager for AI agents on Sui. | Developer experience: the documentation site and the npm releases of @swarmnetwork/vault-sdk. |
| 4Fans | Football fan community with live scores for 125 competitions. | Built solo over 14 months and run it: FastAPI, PostgreSQL, 160+ endpoints, 3,200+ tests. |
Also at Swarm: the draw engine, signed reward-claim flow and operations console behind the Agent License Sale (10,000+ NFT licenses, $10M+ on Sui), and swarm-skills, seven review and audit skills for AI coding agents, where a fresh critic scores every change and the next critic re-checks each fix. It runs as the merge gate on our repositories.
- Wrote the Hacken-audited Solidity for AgentXYZ on Avalanche: the TRADER token and its vesting contract, Hardhat and ethers tooling, through audit to mainnet.
- Audited and hardened an NFT rewards protocol paying USDC to holders: upgradeable reward vault, EIP-712 signed receipts, payout ledger, claim verification.
- Designed token models for 30+ launches at CodeCraft Group and Seedify.
Rewritten in 2026 as tested packages with CI, each with a short report:
- detectingPhishingWebsites, eight classifiers on the UCI phishing dataset
- vertexCover, greedy 2-approximation in C++17 with fuzz tests and benchmarks
- footballDataVisualization, StatsBomb match visualisations, package and CLI
- labelMicroTopicRelationships, active-learning benchmark for prerequisite labelling, graduation project
- wsnGameTheory, game-theoretic jamming simulation in NumPy