A claims desk that closes its own loop
A permissioned agent reads intake documents, drafts adjudication, and escalates anything below a confidence floor. Every action is traced and replayable.
Reference architecture · Applied ML / NLP
A custom pipeline for a 40-year archive
OCR repair, entity resolution and a domain fine-tune turned an unsearchable corpus into a retrieval layer with a defensible accuracy benchmark.
01Applied ML / NLP
The problem
Forty years of documents, much of it poorly scanned, with entity names that changed spelling across decades. Off-the-shelf retrieval returned confident nonsense.
02Applied ML / NLP
What we built
We built a seven-stage pipeline: OCR repair, layout recovery, entity resolution against a canonical registry, chunking tuned to document structure, embedding, reranking, and a domain fine-tune for the last mile.
03Applied ML / NLP
What it proved
The headline number is the honest one: F1 rose from 0.63 on the baseline to 0.91 on the benchmark we built with their archivists.
→What it is held to — illustrative figures
The numbers
Each figure is the measurement the system is held to. These are placeholders from the design comps until client-approved numbers replace them.
Pipeline
Measured continuously; a regression here fails the build.
Retrieval F1
Measured continuously; a regression here fails the build.
Corpus
Measured continuously; a regression here fails the build.
This reference architecture sits under our applied ai & nlp capability. Its figures are illustrative until a client approves the real ones.
↺
Other reference architectures
Market-data plumbing that stops guessing
Normalised multi-venue ingestion with deterministic replay, so teams can reproduce any minute of the day. Infrastructure only — no strategy, no advice, no return claims.
Settlement rails with deterministic tests
Contracts and an indexer with a full replay suite, so every state transition is reproducible before it ever reaches mainnet.
→Start here
Let's build
the impossible.
Tell us the system you can't get built. We come back with a short, paid discovery — a clear plan and a fixed first milestone — usually within two working days.
01
Discovery
A working session to map the problem and define what "good" is measured against.
02
The plan
Architecture, milestones and a fixed first deliverable — yours to keep, either way.
03
We build
Embedded with your team or as a dedicated pod, shipping with traces, evals and docs.