Stas B.

I embed with your team, run the discovery calls, and take an AI workflow from "can this be automated?" to a system running in production. Recent: the AI layer behind three live e-commerce brands, 2M+ generations served, about 600 orders a day. Forward deployed AI engineer, technical lead or fractional CTO for AI products: production AI agents, RAG, evals and guardrails, plus audits and takeovers of AI-built code. 7 years of production software, most of it on systems that cost money when they break. Top Rated Plus, 100% Job Success. HOW AN ENGAGEMENT RUNS 1. Discovery: a black-box audit of your live app, a risk register and a line-item estimate before any contract. 2. Measure first: I run the candidate pipeline on cases your team already handled by hand and show the error rate by case type before anything is automated. 3. Build and ship: agents, retrieval, evals in CI, tracing and cost dashboards, deployed and monitored. 4. Handover: docs, runbooks and a walkthrough so your engineers own it. WHAT I HAVE SHIPPED - AI layer of three AI-portrait e-commerce brands: 2M+ generations, a multi-provider fallback chain (Gemini, OpenAI, Anthropic, OpenRouter) with key rotation, an LLM judge, prompt and sampling A/B tests. - Measured before trusting: backtested a paid vision classifier on 34,779 production generations, found it contradicted itself 21.9% of the time, and replaced it with a local ONNX pipeline (about 50 ms instead of 2.6-3.6 s per photo, no per-call cost). - A LangGraph multi-agent service (qualifier, finder and outreach agents) and an MCP server exposing 65 tools over OAuth 2.0, 18 months in production. - A credit ledger where every model call reserves budget before the provider is touched. Zero unreserved calls, enforced in code. - Rebuilt an AI exam-prep SaaS GCP-native in 8 weeks: eval harness gated on agreement with human labels, about 65% lesson cost cut, then a paid audit with 20 verified findings. - RAG on real documents: chunking, embeddings, pgvector retrieval, prompt-injection scanning on the ingest path. - Before AI: payments orchestration across 100+ providers for an iGaming PSP aggregator, and real-time drilling dashboards used by engineering teams at ExxonMobil, Baker Hughes, Halliburton and Nabors. HOW TO START - AI agent or codebase audit: fixed price from $1,500, 5 days. Where it fails, ranked by what it costs you, with a fix plan any engineer can execute. - Eval harness and guardrails: from $2,900, 1 to 2 weeks. - Build or embedded FDE work: $85/hr, 30+ hrs a week. - Consultation: $150 for 60 minutes. WORKS BEST WHEN - something is in production, or about to be - one person on your side can decide - you want the failure mode named before you pay for the fix Kyiv, remote B2B contractor. New York mornings covered daily (9:00-13:00 ET), full CET overlap. Stack: Python, FastAPI, Django, LangGraph, LangChain, Claude, OpenAI, Gemini, MCP, Langfuse, Postgres and pgvector, Supabase, Redis, Celery, TypeScript, Next.js, React, Node, AWS, GCP, Cloud Run, Vertex AI, Cloudflare.