A useful review router with an uncalibrated confidence score
Document extraction with confidence routing, measured on CORD receipts: the auto/review split works while the score is not monotonic with correctness.
Work
Work
Worked examples of the problems we take on and how we approach them.
A construction document pipeline, measured at every handoff
Tender notices and construction PDFs through intake, extraction, routing and output, with the loss measured at each step, not just the model.
What a RAG query costs in tokens, measured not estimated
Token accounting for a local RAG pipeline: what k and chunk size cost, where prompt tokens go, and why we refused to publish latency.
We measured what a stock PII redactor leaves behind
Stock Microsoft Presidio, measured on real court text and synthetic PII: per-entity leak rates, the partial-redaction trap, and threshold collapse.
We broke a default RAG stack on purpose
An audit of a default-configured RAG stack: which failure modes trigger, which do not, and what each configuration change bought.
An MCP server over internal systems
An illustrative example of MCP server development: a narrow, permission-aware tool surface over internal systems, for the AI clients a team uses.
Invoice exception handling, end to end
An illustrative agentic workflow automation pattern: an agent that handles the invoices a rules-based pipeline rejects, and escalates the rest.
Tell us what the work looks like
Describe the process you want off your team plate. You will get a straight answer on whether agents are the right tool for it, including when they are not.