Dmytro O.

Ex-DataRobot engineer, with technical product roles at DraftKings, Sift and Hypatos. 12 years in production systems, now building LLM applications: hybrid search over 2M+ CVs, a compliance agent on a public surface, eval harnesses for AI extraction.

What I build is the layer around the model: evaluation suites that catch regressions before users do, guardrails and structured outputs so the model can't go off-script, human-in-the-loop where a decision carries legal or financial weight, and prompt-injection defence for anything user-facing. I've built it for startups and scale-ups in the US, UK and Israel.

Recent projects:

• AI search API over 2M+ CVs across 14 brands, hybrid vector + semantic ranking (UK recruitment platform)
• Employment-classification engine where a deterministic rules engine stays the source of truth and an LLM explains its decisions, with a conversational AI assistant in front (UK recruitment, compliance-critical)
• Text-to-SQL analytics agent on Azure with a medallion lakehouse underneath (US logistics)
• LLM validation framework and QA tooling for an AI lease-extraction platform
• 100+ ETL pipelines re-engineered, infrastructure cost cut 10x (Netherlands energy)

Underneath every LLM system I ship is a data platform I could have built myself - and often did. Lakehouses on Databricks and Azure, 100+ pipelines on Airflow, Snowflake and BigQuery warehouses for fintech and telecom. If you need the platform as well as the model layer, that's one engagement, not two.

Typical engagements:

• Evaluation harness and guardrails for an AI feature you already have; fixed scope, 2-4 weeks
• RAG or agent system from prototype to production
• Data platform to support an AI product (Databricks, Snowflake, BigQuery, Azure)
• High-performance API or backend for an AI product: search, retrieval, classification at scale