I build backend services, AI pipelines and internal tools: FastAPI and PostgreSQL on the service side, LLM agents where they earn their place, Docker Compose for everything that has to run somewhere.

Core stack: Python · FastAPI · PostgreSQL · SQLAlchemy · Docker · LLM APIs · integrations · pytest

mtbank-ai-call-analytics — call analytics service: local ASR and diarization, four bounded LLM agents, deterministic scoring and grounding, OpenWebUI Pipeline plus REST, 577 offline tests.

crypto-payment-gateway — payment service: invoices and hosted checkout, merchant API and Python SDK, signed webhooks with retries, deposit-address leases, chain confirmations, sweeping, and an append-only ledger that the runtime database role cannot alter.

balance-tracker — multi-tenant crypto portfolio tracker: exchange balances via CCXT and on-chain wallets, integrity checks before any write, a refresh pipeline with caching and circuit breakers, REST API plus a Telegram bot with plan limits. 280 tests run offline.

tg-radar — index of public Telegram channels: keyword and graph discovery, t.me/s crawling, PostgreSQL state, Vespa hybrid retrieval (BM25 + HNSW) with a PostgreSQL full-text fallback, FastAPI surface for agents.

rag-tender-sql — a RAG service that runs without API keys (local embeddings, extractive answers with citations) next to a PostgreSQL tender-platform schema where integrity rules and analytical SQL live in the database.

• Fail-closed contracts: a model response that does not match its schema should break the run, not silently degrade it.
• Deterministic checks around probabilistic parts: scores, money and permissions are computed by code, not by an LLM.
• Reproducible local runs: each project above starts or demos without my credentials.