Technology Intelligence System
Hybrid retrieval and query orchestration for technical, regulatory and market documents across several business teams in a sensitive industrial setting.
- Role
- Lead AI engineer
- Organization
- Large European aerospace and defence organization
- Context
- Professional
- Capabilities
- Knowledge architecture, Document ingestion, Hybrid retrieval, Structured retrieval, Vector search, Reranking, Query routing, Evaluation, Cloud deployment
- Technologies
- Python, FastAPI, SQL, Vector database
Overview
A retrieval system for technical intelligence: people needed to ask questions across mixed document stores — technical, regulatory, market — without dumping everything into a single embedding index and hoping.
I led the AI engineering: scoping, knowledge design, hybrid retrieval, query routing, evaluation, and the path onto the client’s infrastructure. The consulting team and client stakeholders owned the surrounding delivery. Client operations and all source-document production were not mine.
Problem
Business teams were exploring a large, sensitive corpus with tools that did not combine structured filters, semantic search and follow-up reasoning in a way they could trust. The useful question was rarely “search harder”; it was “which store, which filter, which model, and what should we refuse to answer.”
Context and constraints
The setting is a large European aerospace and defence organisation. That band is as specific as this page is allowed to be. Exact dates, unique volumes, screenshots and security architecture stay off the page.
Access, classification and audit constraints shaped the design more than model choice. A system that retrieved the right paragraph but could not explain how was not usable.
Approach
The architecture combined:
- metadata and knowledge modelling so filters were first-class, not an afterthought
- SQL-backed structured retrieval next to vector search
- reranking and query routing, including when to refuse or ask a clarifying question
- lightweight agent steps only where a single retrieval call was not enough
- evaluation as a loop, not a slide
I do not claim a particular document count, latency, or user number here. Those figures are not published.
Tradeoffs and limitations
Hybrid retrieval is slower to design than “embed everything.” Routing adds failure modes. Evaluation on a classified corpus is awkward: the interesting error cases cannot be shown. The page therefore describes a capability, not a benchmark.
Confidentiality
Client name, logo, internal systems and unique metrics are withheld.