Services

Five practices. Each one has shipped.

We take on work in these five areas. Most engagements touch two or three of them, because problems rarely respect the categories.

Agentic AI

Agents that hold a tool belt, know when to stop, and can be audited afterwards.

What we build

Multi-tool agents wired to your real systems: CRM, ledger, ticketing, warehouse.
Orchestration across steps — planning, retries, human approval gates on anything irreversible.
Retrieval over your own corpus with citations, not a general chatbot with a company name on it.
Guardrails: allow-lists, spend caps, refusal policies, and an eval suite that runs on every prompt change.

An example

An accounts-payable agent for a distribution group. It reads the invoice PDF, extracts line items, matches them to the purchase order, applies the variance policy, posts the journal entry, and pings finance only when something disagrees.

Stack

TypeScript
Vercel AI SDK
Gemini / GPT
pgvector
Temporal
Postgres
OpenTelemetry

Proof point

2,400 invoices/month · 91% straight-through · 4 min average human touch on the rest.

ML / Data Science

Models trained on your history, evaluated against the decision they actually inform.

What we build

Demand and revenue forecasting with backtests you can read.
Propensity, churn and risk scoring wired into the tools your team already opens.
Custom model training and fine-tuning where an off-the-shelf API genuinely isn't enough.
Monitoring for drift, plus a retrain schedule that someone owns.

An example

Weekly demand forecast for a 4,100-SKU catalogue with seasonal and promo effects. Output lands as a reorder list, not a notebook, with stock-out alerts pushed into Slack.

Stack

Python
PyTorch
scikit-learn
LightGBM
MLflow
Airflow
DuckDB

Proof point

Forecast error cut from 31% to 12.4% MAPE. Stock-outs down 44% in two quarters.

Data Analysis

One number, one definition, one place to find it.

What we build

Warehouse modelling: raw sources to tested, documented marts.
Pipelines with alerting, so a silent failure isn't discovered at a board meeting.
Dashboards built around the four decisions a team actually makes each week.
Ad-hoc analysis when the question is one-off — we won't sell you a platform for it.

An example

A margin-by-channel model for a multi-brand retailer, reconciling Shopify, Amazon and wholesale into a single revenue definition that finance and growth both signed off on.

Stack

dbt
BigQuery
Postgres
Dagster
Metabase
Evidence
SQL

Proof point

Month-end close shortened by 6 days. Three competing revenue reports retired.

Web Development

Product engineering. Type-safe, tested, and fast on a bad connection.

What we build

Customer-facing products and internal tools, from first commit to production.
API and integration work: webhooks, auth, third-party systems that document themselves badly.
Performance and accessibility as build criteria, not a cleanup ticket.
CI, preview environments, and infrastructure you can hand to another team.

An example

A claims portal for an insurance broker: document upload, status tracking, role-based access for 40 staff and 900 clients, replacing a shared mailbox and a spreadsheet.

Stack

React
TypeScript
TanStack Start
Tailwind
Postgres
Cloudflare
Vitest

Proof point

LCP 0.9s on 4G. WCAG 2.2 AA audit passed with no critical findings.

E-commerce

The storefront is the easy half. We do the other half too.

What we build

Headless storefronts and custom checkout flows where the default one costs you money.
Catalogue, pricing and inventory sync across marketplaces and the warehouse.
Post-purchase automation: delivery notifications, returns, review requests, fraud rules.
Analytics that tie spend to contribution margin rather than last-click revenue.

An example

A bundle-and-subscription checkout for a supplements brand, plus an order-status SMS flow that cut 'where is my parcel' tickets by more than half.

Stack

Shopify Hydrogen
Medusa
Stripe
Twilio
Klaviyo
Postgres
Next.js

Proof point

Checkout completion up 9.1 points. WISMO tickets down 57% within six weeks.