Messy biology in. Clean sequence out.

Matching on skills, not job titles.

Most job boards just search the text of a posting. This one parses your CV against an ontology of 67 canonical biotech skills (AAV, oncolytic vectors, RT-qPCR, Nanopore, FACS, UF/DF, cGMP) and builds a weighted profile of what you can actually do at a bench.

Every posting gets scored against that profile. It surfaces roles a keyword search would bury, and it shows you the gaps too: skills a role wants that your CV doesn’t evidence yet.

There is no backend. The PDF is parsed by JavaScript in your own browser, and the postings come from the public ATS APIs companies already serve. Nothing is stored and nothing is sent anywhere, including to me.

Every volume is derived, not asserted.

The model proposes concentrations; a deterministic engine derives the volumes from C₁V₁ = C₂V₂. When the two disagree you get the derivation, not a silently corrected number. Overflow is an error rather than a quietly redefined reaction volume, and a volume below what anyone can actually pipette comes back with an intermediate-dilution recipe instead of a shrug.

Behind it sits a repository of 620 commercial kit products across 16 vendors: NEB, Thermo, MilliporeSigma, Promega, QIAGEN, Roche, Bio-Rad, Takara, Agilent, Illumina, Zymo, 10x, IDT, Beckman, PacBio, Oxford Nanopore. Every catalog number was confirmed on a fetched vendor page, and anything I couldn’t confirm was left out rather than guessed at. Name a kit and the SOP gets written around that kit’s real chemistry.

References are checked against Crossref and PubMed and come back in one of four honest states: verified, mismatch (a real DOI attached to the wrong paper, which is the hallucination signature), not found, or unchecked when the registry itself is unreachable. Unreachable isn’t the same as nonexistent, and the badge says which one it is.

Nothing is stored on a server. Protocols, immutable versions, hash-bound signatures and an append-only audit log live in your own browser, and the Word, Excel and robot worklist files are built there too. Across a full test session (save, sign, reload, export, encrypt a backup, wipe, restore) the page made exactly one server request, and it was the health check.

A protocol scored 99.2 and PASS. It was unrunnable.

The scaffold I started from called itself an ISO/GLP compliance system. Its accuracy score was clamped in code to the range 99.0-99.9, which meant it could not fail, and it printed compliance language for checks that were never written. Then it handed me a 24-plex direct RNA nanopore protocol at 99.2 out of 100, green PASS, carrying four independently fatal defects.

To be fair to the checker, its own disclaimer said it verified arithmetic, structure and citation resolution, not science. But a 99.2 next to a green PASS reads as an endorsement, and people will run what you hand them.

So the score got replaced by an auditor that can return FAIL, that shows coverage beside every score, that names what each dimension didn’t look at, and that makes no compliance claim anywhere. Nine platform-chemistry rules now carry published vendor constraints, each firing only when the document names the platform, each citing the constraint behind it. They’re prohibition-aware too: a protocol that says “do not shear” is following the rule, not breaking it. That distinction cost me a round of false positives to learn.

What it still won’t do is tell you whether the science is right for your question. It checks arithmetic, structure, citation resolvability and known platform chemistry, and it says so on screen and on every export. That’s a floor, not a reviewer.

Nine checks, and every one of them is arithmetic.

No model, nothing inferred. Date subtraction, medians and counts. Each check shows the number it fired on, so you can look at that number and disagree with it.

The one I care about most is invoice aging, because it measures a venue against its own settled history rather than a flat 30 days. A venue that has always paid in 45 isn’t late on day 31. A venue that has always paid in 10 is.

It earned its keep the week I finished it. It flagged a night billed at a small fraction of what that venue paid on every other night of the year, with a host cut of the same odd amount sitting on the same date. The event fee had been filled in from the host-cut column.

Two bugs surfaced while I was building the outlier check, and both were the same shape. A venue whose fee never varies has a median absolute deviation of zero, so the check silently never fired on it. And the app announced that a venue’s fee “has never varied” any time that statistic came back zero, which is true whenever more than half the nights simply match. Both now read the actual values.

It told me my own estimated tax was wrong.

The app had always counted a night’s fee on the night it happened. That is the accrual basis, and I had never labelled it as such, which meant the estimated tax figure was quietly including money nobody had collected yet.

Both bases now sit side by side, and the gap between them is broken out rather than asserted: invoiced but not yet collected, hosted but never invoiced, and a timing line for invoices dated in a different year from the night they cover. That third line exists because the first two don’t account for the whole difference, and a decomposition that quietly fails to add up is worse than no decomposition at all.

The reports screen translates the books into double entry, every record becoming a balanced pair. Balancing proves very little on its own, since a translation that halved every figure would balance perfectly, so it is also tied back to the totals the app already asserts elsewhere. If any of those ever disagree, the screen says so instead of showing a tidy number.

Mileage is deliberately absent from the trial balance. It is a tax deduction, not a movement of cash, and including it would make these numbers disagree with the bank. It stays on the Schedule C export, where it belongs.

Banks disagree about what a purchase looks like.

Chase writes purchases as negative numbers. Amex writes them as positive, which is the exact opposite. Capital One splits debit and credit into separate columns, Citi does the same and adds a status column, Wells Fargo ships no header row at all, Discover keeps the transaction date and the posting date apart, and European exports arrive with semicolons, day-first dates and 1.234,56 where the decimal should be.

All of those import without configuration, because the layout is read off the file rather than declared. Amounts are the one thing the app refuses to be quiet about, since getting the sign backwards inverts every number downstream. It tells you what it decided, in a sentence with the totals in it, and puts a Flip button next to the sentence.

Re-importing the same file is safe. Rows match on date, amount and merchant, so a second import of the same statement adds nothing. It counts rather than de-duplicates, which is the distinction that matters: two genuinely separate $4.75 coffees on the same afternoon both survive. Every import undoes in one click.

The interesting part is what stopped charging you.

Subscription detection is easy to do badly. Group by merchant, call anything monthly a subscription, and you have built something that finds Netflix. Everyone already knows about Netflix.

What is worth surfacing is change: the streaming service that went up 16%, the internet bill that went up 19%, the six entertainment subscriptions that each look reasonable on their own and together cost more than the internet does. The demo shows all three, because the sample generator plants them deliberately, along with a gym that quietly stopped charging and was never cancelled.

The Plaid integration is complete: link flow, token exchange, /transactions/sync with cursor pagination, and pending-charge handling. It ships switched off, because Plaid keys belong to a person rather than to an application. Sandbox mode uses fake banks, so the whole flow can be driven end to end before it is ever pointed at a real one.