All work

AI retrieval2025 — present

Nexus AI Workspace

StackWise · Enterprise document intelligence

Nexus AI Workspace — screenshot of the live site

<500ms

p95 retrieval

1000+

Pages parsed / hr

100%

Traceable citations

The problem

Professional teams sit on large document sets — contracts, SOPs, policies, technical specs — that keyword search cannot answer questions about. Generic LLM tools answer confidently but cannot show where the answer came from, which rules them out for regulated work.

What I did

  1. 01Ingest: Docling parses PDFs, contracts and reports — including complex tables, headers and multi-column layouts — into clean structured chunks.
  2. 02Index: chunks are dual-indexed, as BM25 keyword tokens in OpenSearch and as sentence-transformer dense vectors, so lexical and semantic matches both hit.
  3. 03Reason: a LangGraph agent grades retrieved passages and decides whether to fetch more context or rewrite the question before answering.
  4. 04Answer: responses carry exact page numbers, quoted excerpts and a confidence score, so a reviewer can verify every claim.

Outcome

Every module ships containerised and independently versioned against a shared knowledge engine, with controls aligned to ISO 27001, SOC 2 principles, AES-256 encryption and GDPR/HIPAA privacy requirements.

Stack

PythonLangGraphOpenSearchDoclingSentence Transformers

Senior Software Engineer · StackWise