Work

The problems, and what actually moved

Clients are not named. Sector and size are enough to judge whether a problem resembles yours, and the figures below are the ones that were measured rather than the ones that sound best.

Aerospace machining

Turning five years of quotes into a price anchor

Part matching
Nearest neighbours
Browser-side

Try the tool

The problem: Estimators priced new enquiries by searching old quotes from memory. Nothing was reused systematically, so similar parts left the building at inconsistent prices and every enquiry cost hours.
What we did: Matched each new part against historical ones on attributes known at quote time — material family, envelope, tolerance, feature count, quantity — rather than on cost fields a new enquiry does not have. Weights were tuned against held-out parts, and the engine refuses to answer when nothing close enough exists.
Where it landed: A public, self-serve version runs entirely in the browser, so a shop can try it on its own history without the file leaving the machine.

Recall at ten on 1,000 held-out parts. Tuning the feature weights is worth 12.8 points; parsing messy material and dimension strings before matching is worth 22.6 more.

95%
recall at ten on held-out parts: ±36.5%
median price error: <1s
per enquiry

Waste and recycling

A collection business that was not losing money after all

Financial modelling
Unit economics
Break-even
Route optimisation

The problem: The operating plan showed a loss and the conclusion was that the contract was unviable. The plan rested on three assumptions that had never been checked against the vehicles actually in the yard.
What we did: Rebuilt the model from gross vehicle weight upward. Payload had been understated by nearly a third, which cascaded into route counts, fuel, driver hours and disposal trips. Re-optimised the rounds against the corrected payload and re-derived the subcontract rate.
Where it landed: The corrected model turned a projected loss into a profit at a defensible rate, and answered the questions that followed from it: break-even per vehicle, what to pay a subcontractor, and which sites were worth keeping. Every assumption is traceable and adjustable.

Rounds needed a year against payload per truck. The plan assumed 1.5t; the vehicles carry 2.2t. Everything downstream — fuel, driver hours, disposal trips — moves with this one number.

31%
fewer collection rounds a year: 3
assumptions driving the whole gap: Loss → profit
on the same fleet

Global online marketplace

Knowing what an auction placement is actually worth

Bidding models
Auction data
Expected value

The problem: When placement is sold by auction, bidding low puts you under the fold on a page the customer had already decided to buy from, and bidding high means paying retail for demand that was coming anyway. What a placement is worth varies by listing, by market and by device — none of which is visible from the bid side.
What we did: Matched auction outcomes back to what happened afterwards, at the level of the individual listing rather than the campaign: the position won, the price shown beside the competition, and whether it converted. Bids then follow expected value rather than a flat rule, which is the only way one budget behaves differently in a market you are already winning than in one you are not.
Where it landed: Bidding logic and the reporting behind it, running across several placement channels at once.

Listing
level bidding, not campaign level: Expected
value, not a flat rule: Multi
channel, bid simultaneously

Consumer marketplace — our own product

parkingnetherlands.com: a live mobility site priced on public data

Product
Open data
Bidding models
SEO

parkingnetherlands.com

The problem: Drivers and EV owners cannot easily find what a stop actually costs: the electricity, the parking underneath it, and the cheaper option a short walk away are quoted in three different places, if at all.
What we did: Built and run a consumer site over national open data — every public charge point and register-listed garage, priced for a specific stop. Revenue comes from affiliate and auction-based placements, which means modelling what a click is worth by placement and geography rather than bidding flat.
Where it landed: parkingnetherlands.com is live, indexed and growing — the clearest proof we ship and operate rather than only advise. The tariff chart below is its own data: median first-hour price across the fourteen cities it covers.

Median first-hour tariff across the fourteen cities the site covers. Amsterdam is not the expensive one — a result that only shows up once every garage tariff is in one place.

196k
public charge points, live status: 323
garages and P+R sites with official tariffs: 14
cities compared

Hardware manufacturing

Telling demand signal apart from reorder noise

Forecasting
Inventory
Supply chain

The problem: Planning ran on MRP output nobody fully trusted. Warehouses held stock that did not reflect demand, and the gap between what the system said and what planners did was widening.
What we did: Separated genuine demand from replenishment artefacts at component level, sized safety stock against measured variability instead of a flat rule, and made the forecast explain itself so planners could see why a number moved.
Where it landed: A component-level view across the warehouse network that planners could challenge, and a clear account of where MRP alone was never going to be enough.

Component
level forecasting, not category level: Multi-site
warehouse network: Variability
based safety stock

Internal tooling

The hour a day nobody had budgeted for

Internal tools
ERP
Browser extension

The problem: A finance team exported from its ERP daily. The export was a legacy format that modern spreadsheets open with a warning and mangle on save, so every report began with the same manual repair.
What we did: A browser extension that intercepts the download and converts it in place, plus a query console and a parser for the same platform. None of it is clever. All of it removes a step that was costing real time every single day.
Where it landed: The repair step disappeared. The tools are the kind that never reach a roadmap because no one logs them as a problem.

Daily
manual repair step removed: 3
tools across one ERP platform: In-place
conversion, no workflow change

B2B growth

A qualified pipeline built from public sources

Data acquisition
Enrichment
Go-to-market

The problem: Outbound was guesswork. There was no defensible list of who was worth contacting, so effort went wherever attention happened to land.
What we did: Built a crawler and enrichment pipeline over public registries and company data, with deliberate rate limits and a qualification model that scored fit rather than just collecting rows.
Where it landed: Thousands of qualified leads with the reasoning attached, and a repeatable pipeline rather than a one-off scrape.

5,792
qualified leads delivered: Public
sources only: Repeatable
not a one-off scrape

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.

Start a conversation