Yevhen Herasimov
AI engineer building agentic systems — LLMs, RAG, and multi-agent architectures that ship to production rather than stopping at the demo.
Solution architecture is most of the work: which tools an agent is given, how retrieval is shaped, where state lives, what a run costs. The hard parts are rarely the model — they're the scaffolding around it, and what the system does on the turn it gets something wrong.
Google Cloud certified Professional Machine Learning Engineer.
- Architected and shipped a serverless agentic Copilot on AWS — Bedrock Agents, Lambda, API Gateway, RDS and Transcribe, deployed via SAM — putting HRIS data and the platform itself behind a conversational interface.
- Owned a quarterly and annual time-series forecasting platform end to end on SageMaker Pipelines.
- Took three businesses from AI curiosity to shipped integrations, with measurable gains in operational efficiency and key metrics.
- Designed end-to-end LLM solutions on AWS, and made the ROI case for them directly to business leaders.
- Led the architecture of a medical chatbot on GCP with Dialogflow CX, hitting 95% accuracy matching user intent against an internal catalogue of medical services.
- Trained a state-of-the-art model generating realistic lip movement from audio on AVSpeech, and cut its training time 3× by refactoring the pipeline.
- Drove constrained code generation with GPT models and LangChain for LLM-generated websites.
- Built an end-to-end breast-cancer staging pipeline — a CNN scoring individual patches feeding regression models for the final call — reaching 0.635 MSE on the holdout set against a 0.522 best.
- Wrote the patch-extraction pipelines for Whole Slide Images and made data loading and training 10× faster.
- Also served as Program Ambassador in Ukraine, following the Applied Data Institute cohort.
- Built a Flask REST API over MySQL, using raw SQL and ORM where each fit, with auto-generated Swagger docs.
- Integrated real-time chat, video calls, a collaborative document editor and remote storage; containerised the apps with Docker.
Numbers before opinions. I don't recommend a change I can't attach a dollar figure to.
Instrument, then fix. You don't have observability until you've used it during an incident. Most engagements start there.
Honest about what failed. Everything I publish includes what it cost me to learn it.
Currently taking engagements.
Reliability & cost audits, embedded work, and advisory.