About
RadVector API answers one kind of question, precisely: given what was seen on an imaging study, what does current, physician-vetted practice say about it — and what is the evidence for that answer? It uses physician-vetted medical data to reach deterministic and NLP-based inference, and it shows its work on every call.
RadVector API is built by iTech Clinical Intelligence, legally iTECH LLC — a physician-led clinical software company. The criteria are curated and validated by a practising, board-certified radiologist, not assembled by a scraper and not inferred from a model's training data.
The same engine already runs in production. RadVector Reporting, our radiology
reporting platform, calls it in-process on live clinical studies every working day. The API is not a
research prototype packaged as a product; it is the decision layer of a working reporting system,
given its own front door. A test pins the in-process answer and the /v1 answer identical,
so the two can never drift apart.
It is a property you can inspect on the response, not a claim in a brochure.
Each body of guidance — pulmonary nodule follow-up, breast assessment, bone density and the rest — is a validated pack with a version. A verdict cites the pack and the version that produced it.
Every pack records who validated it and when. A pack that has not been through that review answers as advisory and is marked not usable for output. There is no third state where an unvetted rule quietly drives a clinical answer.
Published guidance and newly published evidence are returned as two separate streams. A paper from last month informs the clinician without ever being mistaken for the standard of care.
The two halves do different jobs, and the boundary between them is the point.
Given a structured finding — a solid 7 mm nodule, right upper lobe, single, in a 64-year-old at high risk — the applicable criteria pack is evaluated by code. No model is consulted. The same input returns the same verdict today, next quarter and in an audit two years from now, and the exact text the engine read comes back with the answer so a reviewer can reproduce it by hand.
Most radiology lives in prose, so the API will take free report text and extract findings from it. That normalizer is deterministic too, and it is deliberately unwilling to guess: where a sentence describes several lesions and one measurement, it returns the finding with no measurement rather than a plausible wrong number, because a confidently wrong number is the one failure nothing downstream can detect. Every extraction carries its own confidence, and a verdict can never be more confident than the reading that produced it.
One route drafts an IMPRESSION in prose. It writes around verdicts the deterministic engine has already reached, and a mechanical guard then proves it stayed inside them: every measurement, every category, every diagnosis, every follow-up interval and every content word must already appear in the engine's text or a verdict. A draft that adds one word of its own comes back flagged for review, with the word named. That check is mechanical and closed-vocabulary — never a judgement about whether the prose sounds reasonable.