Yaroslav T.

AI Engineer building LLM agents, RAG pipelines and voice AI that run in production, not in demos. Python, FastAPI, Next.js, React, AWS, n8n. WHAT I BUILD: 🤖 AI agents and workflows Agents that plan, hold memory and run multi-step tasks through tool calling. Built with LangChain and LangGraph on OpenAI, Claude and Gemini APIs, plus MCP servers where the agent needs to reach your own systems. Every tool call is logged, with human approval on anything that moves money. 📚 RAG pipelines and knowledge retrieval Retrieval-augmented generation over your internal data: chunking strategy, embeddings, metadata filtering, hybrid search and reranking. Vector databases including Qdrant, Pinecone and pgvector. Answers come with citations, so your team can verify instead of trusting. Every change is measured against a golden dataset, because tuning retrieval by eye is guesswork. 🎙️ Voice AI and conversational systems Low-latency STT to LLM to TTS pipelines on Twilio, with interruption handling, barge-in and backend-controlled dialogue logic. Built a voice bot running 5,000 calls a day that qualifies customers and returns an offer. Chat agents on WhatsApp, Instagram and Messenger through the Meta Cloud API. 📄 Document processing and data extraction OCR, parsing and structured extraction from PDFs, contracts, invoices and shipping documents. Every field carries a confidence score, and anything below threshold goes to a human review queue instead of silently into your database. The hard part is never the model call, it is output that is 90% right. ⚡ AI workflow automation Zapier, Make and self-hosted n8n connected to real business systems: Slack, Salesforce, HubSpot, Monday, Airtable, Notion, Google Workspace, Shopify and Stripe. OAuth, webhooks, rate limits, retries, idempotency keys and sync state, with error workflows and a dead letter queue, built to survive the day the API changes. ⚙️ Backend engineering and data pipelines Python, FastAPI, Django, Ruby on Rails, Node.js, NestJS and TypeScript. PostgreSQL, Supabase, MongoDB, Redis and Elasticsearch. Real-time and event-driven pipelines, including one processing 2.5 million articles a day and serving an API from 10M+ records in around 100ms. 💻 Frontend React, Next.js App Router and TypeScript. Chat interfaces with streaming, tool states and history. Dashboards, admin panels and reporting UI, mobile-first and bilingual where it matters. ☁️ Cloud, scaling and multi-tenant SaaS AWS including Lambda, ECS, RDS and Bedrock, plus GCP, Azure and Vercel. Docker, Kubernetes, Terraform, CI/CD and monitoring. Multi-tenant architecture with row-level security, roles, permissions and audit logging, deployed in EU regions where GDPR, HIPAA or SOC 2 apply. INDUSTRIES Most of my work sits in four: healthcare, with EHR integrations and access models keeping clinics from ever seeing each other's patient data; real estate and PropTech, with CRMs, rental marketplaces and property management platforms; fintech, including a market data platform re-architected from a buckling MVP; and marketplaces, with verification, matching and Stripe Connect payouts. Beyond those: MarTech, EdTech, HR tech, logistics, legal and compliance and B2B SaaS. Clients come to me most often to take over an existing codebase and get it production-ready, often one built fast with Replit, Lovable or Codex, to scale an MVP that stopped holding, or to move AI from demo into production. SELECTED PROJECTS 🛠️ AI procurement and supply chain platform Operations workspace connecting parts catalogs, suppliers, quotations and shipments, with AI agents extracting line items from supplier quotations and requirement documents. React and NestJS on Azure. ⚖️ Regulatory monitoring and document summarization Consultation records collected automatically into the team's own spreadsheet with deadlines and sources filled in, plus AI summarization of the attached documents they mark as relevant. ⚓ Laytime and demurrage calculation assistant Charter parties run 60 to 70 pages and claims were calculated by hand. The system extracts the governing clauses, separates countable from excluded time and produces a calculation report showing its reasoning. 🎬 AI treatment writer for commercial directors Turns a director's creative direction into a structured, brand-aligned pitch treatment, built so the writing stays theirs rather than coming out generic. HOW I WORK: I look at a requirement and work out the right approach instead of wiring an API and calling it done. If an LLM is the wrong tool for part of your problem, I will say so before you pay for it. Most of my code is now written with coding agents like Claude Code, against a spec and test cases I write first, which is where the speed comes from and where unsupervised agents go wrong. Send me what you are building and what is currently breaking, and I will tell you whether it is a two-week fix or a rebuild.