Alexander Stec

Designed and shipped an agentic LLM system that screens product formulations against regulatory standards across multiple jurisdictions—country, retailer, and brand.

As the sole technical founder, I architected a multi-step agent workflow (LLM classification – rule-based verification – human-in-the-loop checkpoint) with structured outputs and deterministic validation at each stage, so the authoritative logic lives in code rather than in the model.

Every result is citation-grounded—each conclusion links directly to the source standard—making outputs auditable by domain experts, and screening that previously took experts hours per formula now completes in under a minute. The system replaced a manual, error-prone review process in a context where a confident-but-wrong answer is worse than no answer at all.

Built and operated a production retrieval system over a large, domain-specific corpus where missing a relevant result carried real cost.

Off-the-shelf semantic search wasn't enough: domain terminology and exact-match identifiers meant pure dense retrieval silently missed valid results, while keyword search missed conceptual matches. I designed a hybrid pipeline fusing dense retrieval (Pinecone), BM25 sparse, and metadata filters via reciprocal rank fusion, with a streaming LLM postfilter as a final safety net.

Rather than a one-shot build, I iterated against real production failure modes—for example, adding domain-specific synonym expansion after observing valid variants being missed. Continuous online evaluations monitor retrieval quality from live behavior, so regressions surface before users hit them rather than after.

Built an automated pipeline that turns messy, multi-format supplier documents—PDFs, Office files, and images—into structured, source-cited data, replacing a slow manual transcription bottleneck.

The pipeline normalizes mixed inputs to a common format, OCRs them with AWS Textract, classifies each document by type to route it to the right extraction logic, then hierarchically extracts 30+ structured attributes per entity across physical/chemical, functional, clinical, and regulatory categories.

Critically, every extracted value cites the source text it came from, so any figure can be verified at a glance—making the output trustworthy enough to drive downstream decisions rather than requiring re-verification. The system freed domain experts from hours of manual data entry each week.