Platform · Data Refinery
Your data is a mess. That is the normal condition.
Send it in whatever shape it is in. What comes back is one record per real person, where every value can show the exact source that states it — the record every question, profile and conversation downstream is built on.
Every important value keeps its source.
Where this came from
Extracted value
- Fact
- Colonoscopy
- Date
- 11 March 2024
- Source
- Inbound fax · 3 pages · page 2
Scanned — read by AI
The source, with the quote marked
Screening colonoscopy completed 03/11/2024. No polyps identified. Recommend repeat in ten years.
The date and the result both appear inside the quoted sentence. A value that does not appear in its own receipt is never accepted automatically.
Uncertain data does not silently become truth. Anything ambiguous, anything that might be about somebody else, and anything read with less than full confidence goes to a person with the quote attached.
Three kinds of mess, and the third is the expensive one.
Structured, but not standard
Exports where the columns were named by whoever built the report. A title row above the header, a totals row underneath, four date formats in one file.
Written down, not filled in
The note where the clinician actually documented it, rather than the structured field they were supposed to use.
Trapped in a document
The e-faxed colonoscopy and mammogram results that arrive as an image, get filed, and never become a field anyone can query. Nobody is paid to abstract those by hand, so nobody does — and the care counts toward nothing.
However it reaches us, a connector only changes how often data arrives, never what the Refinery can do with it. How data comes in and work goes back.
For technical reviewers: the pipeline, the refusals, and identity
Every value can show the sentence that proves it.
Not a confidence score. The actual quote from the actual document, with a link to the page it sits on.
- 01
The quote must be in the source
Checked as a literal substring of the document text. A receipt that does not verify is never accepted automatically.
- 02
The value must be in the quote
The extracted date and result have to appear inside the sentence that was quoted. Without this, an invented value stapled to a real sentence would pass — which is the hole this closes.
- 03
Anything else goes to a person
A fact that might be about somebody other than the patient, a reading the system was unsure of, a number disagreeing with its own quote: none are discarded and none are accepted. They queue for review, and the dataset ships with them excluded and counted.
A receipt is evidence, not a verdict
It proves the source states this. Whether a scheduled procedure became a completed one is a clinical reading, not a parsing problem — so anything ambiguous, anything that might be about somebody other than the patient, and anything the system read with less than full confidence goes to a person with the quote attached. For a scan or a voicemail the receipt verifies against the transcription, and every value read that way is labeled wherever it appears.
Seven steps, and a person at the two that decide things.
Mapping is confirmed before it commits, and a finished run is a version awaiting your acceptance. Everything between them is deterministic.
- 01Your call
Map
Columns matched to what they actually mean, proposed with a confidence and confirmed by a person before anything commits. Nothing runs on a guess about what a column was.
- 02
Read
Delimiters sniffed, encodings recovered, the real header found beneath the report title. Rows dropped above and below it are counted and reported, never quietly discarded.
- 03
Clean
Dates, phone numbers, codes and units reconciled by fixed rules, with every change logged as before, after, and the rule that fired.
- 04
Mine
Facts pulled out of notes, faxes, scans and voicemails, each carrying the sentence that states it and verified twice before it is accepted.
- 05
Resolve
Records that are the same human collapsed into one person. Conflicts queued for a decision rather than settled by a guess.
- 06
Check
Contradictions across sources surfaced, completeness measured before and after, and evidence recovered from documents counted against the gaps it fills.
- 07Your call
Publish
A reviewed dataset with provenance on every value, exportable or handed to Patient Intelligence. A finished run awaits your acceptance; it never overwrites silently.
What it refuses to clean up.
Most of the value in a cleaning step is knowing where to stop.
- A weight recorded in kilograms is kept and flagged rather than converted, because converting would invent a number the source never wrote.
- A lab result of “greater than 60” stays as it is. That is how the lab reported a capped value, and rewriting it would invent precision nobody claimed.
- A two-digit year is only expanded when the full year was never written anywhere in the source.
- Every change, and every value it could not parse, is written to a log that becomes the quality report you read.
One person, however many records they left behind.
Somebody who has been to five of your facilities has five record numbers and nothing joining them. Until that is resolved, every count you produce is rows rather than people.
Three identifiers agree
Name, date of birth, phone, email. Three matching is enough to call it the same person even when a fourth disagrees — the working rule identity teams already use. A date of birth with the day and month swapped is the same typo, not a second person.
Blank never counts
A field empty on either side is absent: neither a match nor a conflict. Missing data can never inflate a match.
Sex is a veto, never evidence
A disagreement blocks the merge outright. Agreement counts for nothing, because a two-value field coincides far too often — it would merge same-named people in one household.
Record numbers never match across sources
They are local to the system that issued them. Treating them as identifiers is how the wrong record gets pulled.
Nothing is discarded
Where sources disagree the most common value wins and every losing value survives as an alternative carrying its own source. Re-running the same data produces the same people.
What comes back.
Evidence you already had
The people whose structured field was empty and whose care is proven by a document you were already holding — each with the quote, and a link to the page.
A quality report for your data team
Every source, every rule that fired, every conflict, and completeness before and after — measured only over people whose sources carry that field, so merging never looks worse than it was.
The questions, not the guesses
Merges the system would not make alone, values that disagree across sources, and facts that did not verify: a queue with a plain-English reason on each, excluded from the dataset and counted in the report.
One record per person
Provenance on every value. People a source reported as deceased are excluded from every count, and the exclusion is disclosed rather than hidden.
What a record you can check makes possible.
The Refinery is not the product. It is the reason the rest of the product can be trusted — every question, every profile, and every conversation downstream runs on a record whose values can be traced.
- 01
Patient Intelligence
A question about your population is only worth asking of a record where the same person counts once and every value has a source. Cohorts come back as people, with the evidence that put them there.
Ask about your population - 02
The patient profile
The conditions, medications, discharges and documents on a patient's story are the Refinery's output, each opening into the sentence that states it.
One patient, the whole story - 03
Continuous follow-up
A contact opens with this patient's own weight, last result and discharge medication because the record behind it is resolved and current — not with a generic question about their condition.
What runs on the record
It also stands on its own. Plenty of organizations need the record cleaned and have no interest in outreach, and the Refinery is sold and run that way: resolve the data, hand back the dataset, stop there. Publishing into Patient Intelligence is a separate, deliberate step that touches nothing the scheduler can dial.
A warehouse stores what you give it
In the shape you gave it. It will not decide two rows are the same human, it cannot read the fax, and it cannot tell you which sentence proves a value.
A pipeline moves and reshapes
It holds no opinion about identity, has no notion of a receipt, and offers no queue where a person confirms a mapping before it commits.
An exchange finds and fetches
What comes back is documents. Moving a document is not the same as making what is inside it countable, and that gap is the whole job of this layer.
One system, four views
The platform overviewSend us the export you would be embarrassed by.
That is the useful test, and the messier it is the faster you will know. A pilot starts from a file, not from an integration project waiting in somebody's queue.
Talk to our team