Work

Migrations, data platforms, ML models and a lot of AI agents.

Between us the team has re-platformed enterprise analytics estates, built lakehouses and semantic layers, shipped machine learning that drives real planning decisions, and delivered a long list of AI agents. Everything here runs in production.

Discuss a similar project See services

Case studies

Five projects, in detail

Client names are withheld under the confidentiality terms we work to. The problems, the architecture and the decisions are described as they happened.

Case 01 Platform migration

Synapse to Databricks, migrated with row-level parity

The problem

An enterprise analytics estate had outgrown Azure Synapse. Query times were rising, costs were unpredictable, and the finance team had quietly started rebuilding numbers in spreadsheets because they no longer trusted the reports.

What we built

A Databricks lakehouse built alongside the existing estate rather than in place of it. Ingestion was rebuilt on Delta Lake with explicit schema contracts, the transformation logic was ported in dependency order, and Unity Catalog took over access control that had previously lived in a spreadsheet of its own.

How it works

Both platforms ran in parallel for the whole migration. A reconciliation harness compared outputs row by row on every scheduled run and raised a ticket for each mismatch, so the differences were explained rather than averaged away. The business signed off the cutover only once the harness had been clean for a full reporting cycle.

Where it got to

Reporting moved onto a single platform with the lineage and access control that the old estate never had, and the parallel run meant nobody was asked to trust the new numbers on faith.

  • Databricks
  • Delta Lake
  • Unity Catalog
  • PySpark
  • dbt
  • Azure
  • Power BI

Case 02 Machine learning

Demand forecasting that planners actually use

The problem

A retail group planned stock from a mix of last year's figures and experience. Fast-moving lines ran out, slow lines were over-ordered, and every planner had their own private adjustments nobody else could see.

What we built

Forecasting models at SKU and location level, trained on sales history with promotions, seasonality, price changes and calendar effects as explicit features. Planner overrides were made a first-class part of the system: captured, attributed and fed back into evaluation rather than applied invisibly in a spreadsheet.

How it works

Models retrain on a schedule and are scored against a rolling holdout, with accuracy tracked per category so drift surfaces where it happens instead of being hidden in a global average. Forecasts are written back into the planning system the team already used, so adoption did not depend on anyone learning a new tool.

Where it got to

Planners work from a forecast they can interrogate and correct, and the corrections became a measurable signal rather than invisible local knowledge.

  • Python
  • MLflow
  • Databricks
  • Prophet
  • LightGBM
  • Delta Lake
  • Power BI

Case 03 AI agents

A multi-agent system handling first-line customer support

The problem

Support volume was growing faster than headcount. The team's own analysis showed most tickets were variations on a small set of questions, but each still needed a human to read the history, check an order and write a reply.

What we built

A multi-agent system where each agent owns one job — classify the request, retrieve the relevant policy and order history, draft the reply, and check it before it is sent. An explicit escalation path hands anything involving refunds, complaints or unusual sentiment to a person, with the context already gathered.

How it works

Retrieval runs over the help centre, past resolved tickets and the order system, and every answer cites its source so an agent can be audited after the fact. A graded question set built from real historical tickets scores every prompt and model change before it ships. Each tool call is traced and costed.

Where it got to

First-line responses are drafted and checked automatically, with humans spending their time on the cases that genuinely need judgement.

  • LangGraph
  • OpenAI
  • Anthropic
  • pgvector
  • Python
  • Zendesk API
  • FastAPI

Case 04 Voice AI

A voice assistant that answers the help desk line

The problem

A help desk lost most of its calls to hold time. Out of hours there was no answer at all, and the voicemail backlog was handled the next morning by the same people who were already behind.

What we built

A voice assistant that answers the phone, holds a real conversation at human latency, handles interruption and background noise, resolves the common requests end to end, and transfers everything else to a person with a summary already attached.

How it works

Speech-to-speech handling keeps latency inside the window where a conversation still feels natural, which is the difference between a system people use and one they talk over. The assistant is bounded to a defined set of actions it may take unattended; anything outside that set becomes a transfer. Every call is transcribed, scored and reviewable.

Where it got to

The line is answered immediately, including out of hours, and calls arrive at a human already triaged rather than cold.

  • Realtime speech APIs
  • Twilio
  • Python
  • LangGraph
  • PostgreSQL
  • Webhooks

Case 05 Semantics & BI

One definition of every metric, and a chat interface on top

The problem

Two dashboards disagreed about revenue in a board meeting. The underlying cause was four different definitions of the same metric, each correct in its own context and none of them written down anywhere.

What we built

A governed semantic layer holding one definition per metric, with the business owner of each definition named. Every BI tool reads from it, so a number cannot be quietly redefined inside a single report. On top of it, a natural-language interface that answers executive questions against those same definitions.

How it works

The chat interface is deliberately constrained to the semantic layer rather than turned loose on raw tables. It generates queries against governed metrics only, which is what makes the answers reproducible — ask the same question twice and the number is the same, and it matches the dashboard.

Where it got to

Metric definitions became a governed asset with named owners, and executives can ask questions directly without the answer drifting from the reported figure.

  • dbt Semantic Layer
  • Snowflake
  • Power BI
  • Python
  • OpenAI
  • Unity Catalog

Selected work

Every project, filterable by capability.

Between us the team has re-platformed enterprise analytics estates, built lakehouses and semantic layers, shipped machine learning that drives real planning decisions, and delivered a long list of AI agents. Everything here runs in production.

Client names and figures stay confidential unless the client has agreed to publish them. We'll walk you through the architecture and the numbers on any of these on a call, under NDA if you'd prefer.

Recognise your own problem in one of these?

A 30-minute call is enough to tell you whether it's the same shape underneath.