I make data make sense.

Poland publishes its vehicle registry as open data, around 18 million vehicles. I needed one model family out of it. Scanning everything would have taken most of a day, so I wrote a probe script first and proved the API supports a server side make filter. That cut the national census to about half an hour and, more importantly, made the count exact rather than sampled.

The collector is resumable, with a checkpoint per window, and a window only commits once it has been fully paged. Government APIs drop connections, and a census that cannot be resumed is a census you will never finish. Aggregates only: the raw rows never leave the machine, there is no personal data, and every surface that shows a number credits the source.

The UK publishes the result of every MOT inspection it carries out. The catch is that the car I care about, the Corolla E170, was sold in Britain as the Auris, and a UK "Corolla" from 2019 onwards is the next generation entirely. Include it and the sample is quietly poisoned. So the dataset is the platform twin, and it is labelled as a proxy on every single surface that shows it, because a number you have to explain is better than a number that is wrong.

The zips are too large to extract comfortably, and the 2023 archive uses a compression format the Python standard library cannot read at all, so the pipeline streams the CSVs straight out of them. Miles convert to kilometres. Only initial tests count, never retests, because retests would flatter every figure on the page.

The valuation grid is 84 baseline cells, one per year, engine and mileage band, plus 119 researched deductions in zloty, one for every inspection finding at every severity. The research ran as a governed fleet: finders proposed numbers, a matching set of verifiers were told to break those numbers, and a plausibility auditor held back anything that could not be defended.

The part I am most pleased with is not the model, it is the loader. It refuses to half load. Row count, currency, price sanity and severity ordering are all checked inside the transaction, and any violation rolls the whole thing back. A worse fault priced above a milder one is not a rounding error, it is a broken model, and the database will not accept it.

Thirty price bands for common parts, in zloty, quality aftermarket rather than the cheapest thing on the shelf, each one independently re-verified against public Polish retailers before it was allowed to load. Where a genuine Toyota part has a published price, the row carries that too, so an owner can see the actual gap rather than being told which to buy.

The two that did not make it are the interesting ones. One sensor had a band that would have quietly priced the wrong variant of the part, and one exhaust component simply does not exist as a specific aftermarket item for this body shape, so any band would have been a guess dressed as a fact. The decisions are written into the generator rather than applied by hand, which means a re-run reproduces the same result, refusals included.

Everything above exists because it ships inside something real. Cario is a maintenance, inspection and valuation companion for one car, deliberately vertical: one model, one country, measured properly before anything wider. Twenty eight screens, feature complete, camera scanning of the vehicle number with recognition running on the phone rather than a server, and data export and deletion built in from the start rather than bolted on.

Every schema change is a numbered migration, which means the database can be rebuilt from zero and arrive in exactly the state it is in now. That is the same discipline regulated pharma taught me: if you cannot reproduce it, you do not really have it.

I built a backtesting laboratory whose entire product is answers I can trust, then pointed it at eight of my own trading ideas across currencies, crypto, metals and equity indices. Every hypothesis was pre-registered before it touched data: a falsifiable claim, a statement of who is on the other side of the trade losing money, explicit kill criteria, and a fixed budget of attempts. Every run lands in an append only ledger, so the statistics stay honest about how many things were tried.

Two tripwires keep the instrument itself honest. A strategy with no edge must earn nothing, and a strategy that cheats by peeking one bar into the future must be caught loudly. I calibrated in both directions with planted synthetic edges: it never cried wolf across 140 checks, and it reliably finds a real edge once that edge is worth roughly twice what trading costs.

The verdict after eight hypotheses was eight kills, no false positives, and a written reason for every death. That is the skill I am actually selling: knowing when the answer is no, and being able to prove it.