Defence consultancy
Retrieval and claim-level attribution across a contested corpus.
live link:
year >
industry >
Defence advisory
problem >
A senior assessment is signed by a person, not a model. The corpus is open-source and contested: think-tank papers, procurement notices, transcripts, court filings. The hard part is not retrieval, it is attribution. A claim with no traceable source and no stated confidence cannot be defended when the client pushes back mid-readout.
what we did >
pgvector retrieval over Postgres, row-level isolation per case, append-only hash-chained audit log, FastAPI services, Next.js and TypeScript client, DOCX export. The model extracts claim-level statements with citations and a confidence value. The analyst stays the author.
result >
By automating repetitive workflows and administrative tasks, the company significantly reduced the time spent on day-to-day operations. Teams were able to focus on higher-value work instead of manual processes.
.
01
15
distributed sources
.
02
45%
time saved
Tasks that previously required hours of manual effort were completed automatically in the background. This resulted in faster execution, improved productivity, and more time dedicated to growth, customer experience, and strategic initiatives.
Sarah Mitchell
Founder, Interlock Ind.
m
Got some other
questions?
