Oleksandr P.

I build AI that works where mistakes cost money: LLM underwriting agents, fraud detection, and the payment and bank-data integrations around them. 8+ years in software, 4,800+ hours on Upwork at 100% Job Success. Recent work for a US consumer lender (ongoing, long-term): - AI underwriting agent: a multi-agent LLM flow that runs in parallel with the in-house underwriting engine, giving each loan application a second, independent assessment based on bank data, credit and ACH reports. Tool calling, durable state, automatic Azure-to-OpenAI fallback, and evaluation baselines in CI so model or prompt changes can't silently degrade decisions. Handles hundreds of applications a day. - Paystub fraud detection: vision models plus bank-side forensic signals from Plaid, built as a pluggable signal aggregator so new fraud checks plug in without touching the core. Flags suspicious documents before they reach a human underwriter. - Model selection: benchmarked GPT-5.x and Grok models on real underwriting cases to pick the best cost/quality balance. - Payments: ACH processing with state machines for NSF and return codes, partial and split card payments, and loan write-off workflows. - Contact center: custom Twilio Flex plugin for call center agents and an AI reply assistant built into it. How I build AI features: evaluation first, then prompts. Structured outputs, provider fallbacks, and tests in CI, the same standards as any other production code. Day to day I work with Claude Code and Codex, often running several agents in parallel, which lets me move from idea to working prototype fast. Stack: AI: OpenAI API, Azure AI Foundry, LangChain.js, LangGraph, Vercel AI SDK Backend: Laravel/PHP, Node.js/TypeScript, NestJS, Redis queues, Temporal Frontend: Vue/Nuxt, React Integrations: Plaid, Payliance, Stripe, DocuSign, Twilio (Flex, Studio), SendGrid Outside fintech, I take on Laravel and Node.js products that need AI features or heavy integrations: SaaS back ends, workflow automation, real-time apps. Have an AI idea that needs to hold up with real users, or an integration that can't break? Send me a short description and I'll tell you how I'd approach it.