Andrii H.

Web scraping and data extraction at scale, then the document AI on top of it: I turn websites, PDFs, scans and legacy databases into clean, structured data and build the LLM pipelines and RAG systems that use it. Document extraction with validation, entity resolution for healthcare data and financial records, AI agents and MCP servers. Top Rated Plus · 100% Job Success · 2,900+ hours · finance, healthcare, real estate, e-commerce. Team of 5 engineers behind me. I build with AI coding agents (Claude Code) day to day and check their output the way I would a junior engineer's: tests, design notes and a change log on every piece of work. On production systems, model calls are logged with their cost, and output quality is tracked per run rather than assumed. Most AI projects stall at the same wall: the team wants an agent or a RAG system, but the data lives in scattered PDFs, broken APIs, half-scraped websites and legacy databases. I do both halves of that problem, so the AI build does not pause for three months while someone "figures out the data". WEB SCRAPING & DATA EXTRACTION Production scrapers for Amazon, Etsy, Mercari and niche marketplaces, real-estate portals, gaming and sports platforms: Playwright / Scrapy with anti-bot rotation, proxies, scheduling, deduplication and clean datasets into PostgreSQL, Google Sheets or your API. A real-estate data lake with daily refresh across 12 sources; an RSS / news aggregator; 200K+ legacy financial documents parsed into PostgreSQL with full lineage. Most of my long-term clients started with a scraping job. LEGACY DATA & DOCUMENT ARCHIVES Legacy data is a specialty: Access (.mdb), FoxPro/dBase, FileMaker, Excel/VBA systems, scanned archives and PDF-locked records, migrated into modern PostgreSQL with full validation reports and zero silent data loss. RAG, KNOWLEDGE GRAPHS, AI AGENTS & MCP SERVERS Document-intelligence systems where every answer is traceable: semantic index (pgvector) plus a Neo4j knowledge graph of extracted facts, so answers surface documents that share no vocabulary with the question. Every passage carries file, paragraph and the chain of entities behind it. Retrieval is hybrid, semantic search over pgvector combined with keyword search, and where the client needs measured retrieval quality we build an evaluation set and track it. Connectors for SharePoint, DataSite and Outlook, background workers, multi-tenant isolation, on-prem option. AI agents (Claude, OpenAI) and MCP servers that expose your data and APIs to any MCP client. MCP servers in production for a client's internal data and tools, with usage logging and per-call cost tracking. AI assistant for an electronics distributor (BOM cross-referencing, 84% catalog coverage); AI ops assistant for affiliate networks (Affise / Keitaro / Voluum / Tune). HEALTHCARE DATA AGGREGATION & ENTITY RESOLUTION Medical Data Graph for a MedTech due-diligence platform: 7.4M US providers (NPPES), 9.3M Medicare procedure records, 17M Open Payments transactions, 2.3M facility affiliations, 194K clinical trials with 133K investigator roles resolved to named physicians, 1.88M PubMed publications with 3.25M author links, linked by a custom matching engine with confidence scoring (95–100% accuracy on high-confidence tiers) and served through a 9-endpoint FastAPI / PostgreSQL API in 2–3 seconds. Services: NPI matching and HCP data validation, physician / KOL target lists, market analysis, the graph deployed in your cloud. Public data only, no PHI. Healthcare datasets, extraction rules and agent outputs are reviewed by an in-house biomedical expert (DVM, 20+ years in laboratory diagnostics and pharmacology): units, reference ranges, drug and biomarker mapping, and LLM outputs checked against the source before anything reaches your database. The same aggregation and record-linkage engine works for companies, contracts, people, properties. QUANT BACKTESTING & STRATEGY MIGRATION Migrated a legacy trading-signal system (VB.NET / EasyLanguage) to Python with signal-by-signal validation against the original; built corporate-action-adjusted market-data pipelines, backtesting with standard risk metrics, and parameter-optimization tooling with out-of-sample checks. I port EasyLanguage, AFL, MQL, ThinkScript and Excel strategies to Python and prove the port before anyone trades it. HOW I WORK: week 1 scope and edge cases locked; weeks 2–3 something running on your data, not slides; then iteration with evals, guardrails and observability. Agents are engineering, not prompts. STACK: Python, FastAPI, Claude / Anthropic API, MCP, OpenAI, Azure OpenAI, LangChain / LangGraph, pgvector, Qdrant, Neo4j, PostgreSQL, Redis, Airflow, Playwright / Scrapy, pandas / Polars, AWS, Azure, Docker. Send me what you are trying to build. I reply within 24 hours on whether I am the right fit.