Work

What we build, and what it changes.

Live engagements, with the numbers they produced. Client names are withheld — we'll make reference introductions on a call.

Conversation intelligence

Service calls, audited end to end

Consumer durables · contact centre and field service

What we walk into

A contact centre handling a very large volume of calls a month, with quality checked on a small manual sample. Everything outside that sample is unread. Nobody can answer what customers asked for and didn't get, which agents are drifting, or how often a renewal was there for the taking and went unmentioned.

What we build

A pipeline that transcribes every call in the languages it actually arrives in, scores each one against the company's own audit rubric, and in the same pass marks the revenue moments — AMC and warranty renewal, exchange, spares and consumables. Supervisors get per-agent scores; the full transcript archive becomes searchable in plain English.

2 in 3calls carry a genuine revenue opportunity
48%AMC capture on the calls the system flagged
~50%of those opportunities closed with no pitch at all — the customer asked and was answered
100%of calls audited, and searchable in plain English

What changes. Coverage goes from a sample to everything, so agent scores stop being a lottery of which calls got picked. And the audit pass starts paying for itself, because the same listen that checks compliance also finds the renewals nobody logged.

Conversation intelligence · calibration

Scored against your own auditors

EV infrastructure · battery-swap network

What we walk into

An operator who already scored a set of calls by hand, against a written rubric, before we arrived. That is rare and it is the strongest asset a buyer can have — it means model quality can be checked rather than believed.

What we build

The same scoring pipeline, calibrated against that human-scored set so machine and auditor can be compared line by line on the same calls, attribute by attribute. Disagreements are the useful output: they show where the rubric itself is ambiguous, not just where the model is wrong.

200calls hand-scored by their auditors, used as the ground-truth set
Monthlyreconciliation against a fresh shared sample

What changes. Agreement with your auditors is measured on a shared sample every month, rather than asserted once during the sale. When the two drift apart you find out in the next reconciliation, not in a quarterly review.

Operations intelligence

Ads and demand, back in sync

Retail and distribution · multi-channel operations

What we walk into

Orders in one system, inventory in another, spend in a third, and a team reconciling all three in a spreadsheet every week. Marketing therefore spends against last month's picture: the fast movers go out of stock mid-campaign while cash sits in stock nobody is pushing.

What we build

A layer that unifies those systems and encodes the reconciliation rules your team carries in their heads, then puts SKU-level demand forecasting on top of it and reallocates spend across channels and creatives against what is actually going to sell. Anomalies — a channel that stopped reporting, a SKU whose returns spiked, a vendor drifting off agreed rates — go to the person who owns that number.

+28%return on ad spend
−18%blended customer acquisition cost
−30%inventory days
−22%working capital tied up in stock

What changes. Spend follows demand instead of trailing it, and the weekly reconciliation stops being a manual job. Exceptions arrive as a message to a named owner rather than waiting in a dashboard for somebody to go looking.

Workflow automation

Agents that finish the routine work

High-volume inbox and catalogue operations

What we walk into

Tens of thousands of inbound mails a month across pre-sales, complaints and partner queries, handled by a team reading each one from scratch. Elsewhere, the same shape of problem in catalogue work — a designer regenerating the same variant by hand, hundreds of times.

What we build

Agents that classify, prioritise, draft and route — trained on your own reply history so drafts land in your team's voice — resolving the repetitive cases end to end and escalating the rest with context attached. On the catalogue side, variant generation as a batch job, and visual search so a salesperson can find a design from a photograph.

25–30kemails handled a month across pre-sales, complaints and partner queries
10×faster first response — hours down to minutes
100%of actions logged and auditable

What changes. A human stays on every edge case and every action is logged, so the automation can be audited rather than trusted. What goes away is the repetitive middle, which is most of the volume.

Bring us a workflow you think AI should touch

Thirty minutes with a founder. We'll tell you honestly whether it should — and if it shouldn't, what would need to be true first.

Book a 30-minute call