Selected work

Case Studies

A snapshot of projects we've delivered. Names and figures are anonymised until we have client sign-off — happy to walk through specifics on a call.

⚠ TODO_REPLACE: these are sample case studies. Replace with real client work as it ships.

Logistics SME · 120 staff

42hrs

saved per week across dispatch team

From spreadsheets to a self-service dispatch portal

The problem: Operations staff were rebuilding the same spreadsheet every morning to plan routes, juggling driver availability, vehicle capacity, and customer ETAs by hand.

What we built: A lightweight web portal that pulled in driver schedules from the existing HR system, customer orders from the order DB, and produced a daily plan that dispatchers could tweak and publish in two clicks.

Outcome: 42 hours of manual work reclaimed every week. Same-day delivery rate up from 71% to 88% within two months.

Engagement: 7 weeks · Stack: PHP + Postgres + React

Mid-sized accounting firm

3 → 1

scattered systems consolidated into one warehouse

A single source of truth for client reporting

The problem: Partner reporting pulled from Xero, a time-tracking tool, and a CRM — each with its own definition of "client", "engagement", and "revenue". Reports took two days to assemble.

What we built: A BigQuery data warehouse with nightly extracts, a small dbt model unifying client/engagement entities, and Looker Studio dashboards for the partner team.

Outcome: Monthly reports now render in under 30 seconds, with consistent definitions. Partners spotted a previously invisible margin issue inside the first week.

Engagement: 9 weeks · Stack: BigQuery + dbt + Looker Studio

Insurance broker

3.2×

document-classification accuracy vs off-the-shelf LLM

A fine-tuned model that reads policy documents

The problem: Underwriters spent hours each week extracting fields from incoming insurance schedules — formatted by ~30 different insurers.

What we built: Fine-tuned an open-source LLM on a curated set of 4,000 historical documents, deployed via a private endpoint so client data never leaves their cloud.

Outcome: Field-extraction accuracy of 94% vs. 29% from the off-the-shelf baseline. Underwriters now review and approve rather than re-type.

Engagement: 12 weeks · Stack: Llama-3 fine-tune + private GPU endpoint

Could yours be next?

Tell us the bottleneck and we'll tell you, honestly, whether we think we can move the needle and roughly what it would cost.

Start a conversation