puodziukas.dev · remote · available Q3 2026

RAG / Vector Data Engineer

Deterministic retrieval. Precision = 1.0. COBOL source data. Qdrant, running on this site.

Precision = 1.0 on a fixed deterministic test set. Not estimated. Measured, the same discipline that reproduces the public 47,312-record COBOL probe.

reproducible · verified · available Q3 2026

The problem

A RAG system can return a semantically similar result that is simply wrong — cosine similarity is not truth. Enterprise use cases need recall that is measured, not estimated.

What exists, reproducible

the public 47,312-record COBOL probe is the proof that a fixed corpus measures precision exactly. The test set is fixed. The expected outputs are fixed. The system passes or it does not. That is the bar.

Relevant projects

• →Qdrant-backed knowledge snapshot: RAG over a private knowledge base (running on this site)
• →Deterministic retrieval test set: precision measured exactly, not estimated
• →COBOL fixed-width records → embedding pipeline (structured RAG over mainframe data)
• →Chunking strategy optimized for long-form structured document corpus, tuned against the deterministic test set
• →Hybrid BM25 + vector search with reranking layer tuned against deterministic test set
• →COBOL mainframe records as RAG source: fixed-width → chunked → embedded → queryable
• →Context window budget optimizer: retrieve exactly what fits, no truncation, no padding

What you receive in 30 days

1. Chunking strategy benchmarked on your document corpus: precision/recall measured (Week 1)
2. Vector collection live with first semantic queries returning verified results
3. Retrieval quality baseline: deterministic test set, measured precision score

Companies at this intersection

Pinecone · Weaviate · Qdrant · LlamaIndex · SingleStore · Chroma · Zilliz · Vespa

All results under NDA. Representative figures. SHA-256 hash-chained audit trail available on request. · puodziukas.dev