Selected work

Engagements, told the way we run them.

No logo walls. Two builds, told in full: the problem the client brought us, the approach we took, and the outcomes we can measure.

ShareLinkedInX
  1. Clinicians reviewing patient notes on a tablet in a modern clinic

    Healthcare · LLM workflow

    Cassia Health

    01

    Clinical notes that write themselves

    The problem

    Cassia's clinicians were spending more of their day writing than practising: consultation notes, referral letters and coding summaries consumed evenings across a fourteen-clinic network, and locum turnover made the backlog worse. Off-the-shelf dictation tools produced text no reviewer trusted, so doctors re-read and rewrote everything — the automation added work instead of removing it.

    The approach

    1. Spent two weeks embedded in clinic workflows before proposing anything — mapping where notes actually originate, who reviews them, and where errors get caught.
    2. Built a domain-tuned drafting system grounded in each clinic's own terminology and templates, with every draft showing its evidence so a reviewer can check the source in one click.
    3. Shipped behind a strict evaluation harness: clinically reviewed test sets, drift monitoring on live traffic, and a quality gate no draft can pass without meeting the network's own review standard.
    4. Rolled out clinic by clinic with the existing team in the loop, converting sceptics first because they wrote the toughest reviews.

    The outcome

    less documentation time across the network
    62%

    less documentation time across the network

    clinics running the system daily
    14

    clinics running the system daily

    of drafts accepted with minor or no edits
    94%

    of drafts accepted with minor or no edits

    The system doesn't write like a doctor — it writes like our best registrar, and it shows its sources.

    — Clinical operations lead, Cassia Health
    ShareLinkedInX
  2. Logistics operations team reconciling daily ledgers on screens

    Fintech · Agents & evaluation

    Ledgerline

    02

    An analyst that never sleeps

    The problem

    Ledgerline's finance team reconciled daily ledgers across payment providers by hand — a morning ritual that consumed three analysts and still left exceptions aging past the point where root causes were traceable. Growth made the problem compounding: every new payment corridor added surface area faster than headcount could cover it.

    The approach

    1. Started from the reconciliation playbook the analysts had refined over years — encoding their judgement calls, not replacing them.
    2. Built agentic matching over the ledger data with a signed audit trail for every match and exception, so the finance team stays the system of record.
    3. Paired the agents with an evaluation harness replaying months of historical ledgers nightly, measuring match precision against the analysts' own past decisions before anything touched production.
    4. Left the exception queue human-governed by design: the agents prepare, the analysts decide, and every escalation trains the next replay.

    The outcome

    of daily ledgers reconciled without human touch
    98.4%

    of daily ledgers reconciled without human touch

    analysts redeployed from reconciliation to review work
    3

    analysts redeployed from reconciliation to review work

    median time to match a ledger line
    1.2s

    median time to match a ledger line

    We stopped hiring for the backlog and started spending the hours on the exceptions that matter.

    — Head of finance, Ledgerline
    ShareLinkedInX

Your case could be the next one on this page.

Every engagement here started as a short conversation. Tell us what you're building and we'll tell you honestly whether intelligence belongs in it.

Start a project