Capabilities
What we are actually good at
8 areas, and they are more alike than they look. Each one comes down to choosing well under a constraint — which round to run, which part to price from, how much to hold, what to bid, what a site has to clear to wash its face. The domain changes; the shape of the problem does not.
Part matching and should-cost benchmarking
Price a new enquiry from the parts you have already quoted.
Contract manufacturers hold years of quotes and almost never use them. An estimator searches by memory, so the same part gets priced three different ways in a year. Matching a new enquiry against historical parts on physical attributes — material family, envelope, tolerance, finish, quantity — turns that archive into a price anchor, in seconds, with the reasoning shown.
95% recall on held-out parts, ±36.5% median price error
Route and collection optimisation
Fewer runs, the same tonnage.
Collection economics live or die on assumptions nobody re-checks: payload, round length, subcontract rate. Rebuilding the operating model from vehicle class upward — then optimising the rounds against it — usually moves the answer further than any routing software will.
Rebuilt model turned a projected loss into profit on the same fleet
Inventory and demand planning
Stock that reflects demand instead of last year.
An MRP run tells you what the parameters say. It does not tell you whether the parameters are right. The work is separating genuine demand signal from reorder noise, sizing safety stock against measured variability rather than a flat rule, and giving planners a reason to trust the number.
Component-level forecasting across a multi-warehouse network
Pricing and bidding models
Know what a slot is worth before you bid on it.
Auction-based inventory rewards whoever estimates value most accurately, not whoever bids hardest. That means tying auction outcomes back to what happened afterwards — at the level of the individual listing, not the campaign — then setting bids from expected value, with a ceiling that is enforced rather than hoped for.
Auction bidding at scale, and a consumer marketplace of our own
Financial modelling and unit economics
Know which number the whole case actually turns on.
Most business cases are one assumption wearing a spreadsheet. Build the P&L from the physical facts upward — what a unit costs to make, move and dispose of — and the break-even falls out of it rather than being argued towards. Done properly it also tells you what rate to negotiate, which sites to keep, and how wrong an input can be before the answer changes.
A rebuilt operating model that turned a projected loss into a profit
Document extraction and classification
Get the number off page sixty, and show your working.
The figure you need is in a table in a PDF, and it has to land in a category somebody else defined. Doing it once is easy and doing it at volume consistently is not, because the judgement calls stop being visible. We split the job: deterministic code does ingestion, parsing, unit and currency conversion and assembly, a model does the reading and the classification, and rules are written at category level so nothing gets a bespoke carve-out. Every decision keeps the reason and the page it came from.
Rejected findings published alongside accepted ones, with reasons
Data engineering and internal tooling
Pipelines, warehouses, and the small tools that remove daily friction.
Most analytics problems are plumbing problems wearing a hat. Modelled warehouses, tested transformations, scheduled jobs that fail loudly — plus the unglamorous extensions and parsers that save a team an hour a day and never make a roadmap.
Browser extension fixing legacy ERP exports, used daily
Decision support and modelling
A model you can argue with.
Board papers and business cases built so the assumptions are visible and every figure is traceable to a source. If the answer changes when an assumption changes, that should be a slider, not a rebuild.
Operating models, scenario analysis, and the paper that carries them
How engagements usually run
A week to find out
We take your messiest real extract and come back with what is actually in it, what it cannot answer, and whether the thing you want is achievable. Fixed price.
Build the thing
Model, pipeline or tool, delivered in your stack with the assumptions visible. You get the code and the reasoning, not a black box and a retainer.
Hand it over properly
Documentation your team can act on, and enough time working alongside them that the thing survives us leaving.
Fixed prices
DORA register of information
The one line of work with a price list: a pre-submission check of the register, remediation when it comes back from the supervisor, and a quarterly concentration report for the board.
Try it on your own file first
Every tool here runs in your browser and uploads nothing, so you can see what we would see before telling us anything at all. When it is worth a conversation, send the real extract and the question you actually need answered.