KIRILL TSYGANOV

Production AI Agents · Enterprise RAG · Low-Latency Systems

I build AI agents that say "I don't know" instead of making things up.

💬 Direct Chat on Telegram • ⚡ Flagship Systems • 📊 All 27 Projects • 💼 Services & Work with Me

PROOF OF ARCHITECTURE · DETERMINISTIC GUARDRAILS · MEASURED METRICS

Sub-millisecond vector indexing & hybrid search without external container bloat.

• The Problem: Production agent loops degrade in speed when querying heavy cloud vector databases for short-term and conversational memory.
• The Solution: Pure JavaScript in-memory HNSW skip-graph indexer with Float64Array storage, hybrid BM25 lexical search, and zero dependencies.
• Measured Metric: Recall@10 98.5% at N=500 → 78.8% at N=5,000 on default settings (random 64-d vectors, brute-force ground truth); P50 0.3–0.9 ms. Tunable via efSearch (98.0% recall at efSearch=128): see the recall/latency curve in the repo.
• 🔗 Explore Repository & Architecture →

A first-line command gate for AI agents: regex rules, risk policies, dry-run rewrite, tamper-evident audit log.

• The Problem: Autonomous LLM agents with bash/tool access can generate dangerous commands (rm -rf, curl-pipe-sh, data exfiltration) when prompt-injected.
• The Solution: 10 threat rules (rm -rf, .env exfiltration, cloud IMDS, curl|bash, reverse shells…), 4 policy profiles (PARANOID / STRICT_CI / AIRGAPPED / DEVELOPMENT), automatic dry-run rewrite and a SHA-256 hash-chained audit log.
• Measured Metric: 10 rules · 4 policies · 13 tests, including checks for known bypasses (base64, variable indirection) and cryptographic hash chain verification. Zero dependencies. Meant to sit in front of an OS-level sandbox, not replace it.
• 🔗 Explore Repository & Architecture →

Empirical AI evaluation harness & 172-task benchmark suite eliminating subjective vibe-checks.

• The Problem: Prompts and agents are typically evaluated with arbitrary single-number averages without statistical significance testing.
• The Solution: Paired McNemar significance testing ($p < 0.05$), Cohen’s $\kappa = 0.718$ LLM-as-a-Judge calibration, and multi-objective Pareto Frontier analysis (Accuracy vs Latency vs Token Cost) across 6 core agent task families.
• Measured Metric: 172 curated tasks · 477 automated tests (100% pass) · Statistical evidence, not vibes: paired McNemar test + bootstrap CIs (reproduce via pytest -v or python3 simulate_pipeline.py).
• 🔗 View PASSMARK → · View BENCH-SUITE →

I help engineering teams, startups, and businesses build reliable AI systems that run in production without hallucinations, runaway token costs, or data leaks.

1. Autonomous AI Agents & Tool-Calling Workflows (FastAPI, n8n, LangChain, Playwright) — structured data extraction, CRM sync, customer support with human-in-the-loop.
2. Enterprise RAG & Grounded Search (Qdrant, ChromaDB, LlamaIndex) — strict citation verification, closed-book guardrails, sub-second latency.
3. AI Safety, Firewalls & Sandboxing (command-pattern scanning, input sanitization, rate-limiting, tamper-evident audit logging).
4. Custom Backend & Automation Engines (PostgreSQL RLS, Stripe state machines, OpenTelemetry observability).

• Rapid Prototyping: Working MVP with unit tests in 5–7 business days.
• Production Deployment: Full Docker Compose / CI/CD and observability.
• Direct Communication: Daily asynchronous updates + Telegram channel.

👉 Ready to automate your operations? Reach out directly:

• Telegram: @therealfullmetal
• LinkedIn: Kirill Tsyganov
• Email: millyrock2900 [at] gmail.com

Разрабатываю производственные AI-агенты, RAG-системы и комплексную автоматизацию бизнес-процессов под ключ:

• Автономные агенты без галлюцинаций: квалификация лидов в CRM (AmoCRM, HubSpot), умная сортировка входящих писем и счетов, боты поддержки с контролем точности.
• Автоматизация e-commerce и маркетинга: мониторинг цен маркетплейсов (Wildberries / Ozon), кросспостинг контента, сквозная аналитика.
• Индивидуальные бэкенд-решения: интеграции Telegram/WhatsApp, вебхуки, биллинг, базы данных с изоляцией тенантов.

💬 Обсудить задачу или получить консультацию: Написать в Telegram (@therealfullmetal)