Hakim

Kadyrov

AI analytics SaaS that lets non-technical users get instant answers from their business database by asking in plain language - no SQL, no manual reports. Pipeline: a natural-language enrichment layer (period parsing, synonyms, fuzzy entity matching) -> LLM function-calling that picks from a whitelist of 25+ analytics functions (never raw SQL) -> parameterized, per-tenant SQL. Security: prompt-injection guard (30+ patterns), all personal data hashed (SHA-256) before it reaches the LLM, strict multi-tenant isolation. Stack: Python, FastAPI, JWT, streaming ingestion (ijson), SQLite/PostgreSQL, billing webhooks, Telegram bot (aiogram), PWA, Nginx + systemd. Also an air-gapped on-premise build with a local LLM (Ollama). ~15,000 lines of Python, deployed in production.

AI analytics SaaS for service businesses (salons, clinics). The owner connects their CRM and within minutes gets an AI chat, a live KPI dashboard, a TV/BI panel and automatic daily reports. Key engineering: a universal CRM adapter system (abstract base + registry) integrating YClients, Altegio and amoCRM (OAuth 2.0) REST APIs, with real-time webhooks and ETL into an analytical store. Same secure AI pipeline: LLM function-calling -> parameterized, per-tenant SQL -> PII unmasking. Personal data is hashed; CRM tokens are encrypted (Fernet). Hardening: fixed multiple SQL-injection vectors and ran a multi-agent security audit (100+ checks). Stack: Python, FastAPI, SQLite (WAL), httpx, aiogram, pytest, Nginx + systemd.