I started Aether thinking the difficult problem was organizing healthcare information. I eventually realized the harder problem was acquiring it without asking healthcare to first become something it isn't.
When I began building Aether, the problem seemed relatively straightforward. A patient might have years of laboratory reports, prescriptions, scans and other medical records scattered across hospitals, diagnostic centers, doctors, phones and folders. Most of that information existed, but very little of it existed together.
The first version of Aether tried to reconstruct the patient from those fragments. Upload a report. Extract the clinically relevant information. Standardize it. Put it against everything that had come before. Begin creating a longitudinal medical history rather than another folder of documents.
Fairly quickly, it became obvious that a timeline alone was not enough.
Health is not simply a sequence of observations. It is a set of relationships across time. A glucose value matters partly because of the values that preceded it. A prescription matters in relation to the symptoms, diagnosis and laboratory results around it. A hospitalization can change the meaning of what happened immediately before and after it. Two abnormalities that seem unrelated in separate reports may become significant when they repeatedly occur together.
That realization led Aether toward the health graph. Instead of treating a patient's history as a collection of documents, Aether began representing reports, parameters, prescriptions, encounters, imaging, vitals, diagnoses and health events as parts of a connected longitudinal structure.
The graph gave the patient memory. More importantly, the relationships inside it began to create context. For a while, I thought that was the hard problem.
It wasn't.
The graph can only know what enters it
The more time I spent inside healthcare workflows, the more obvious another problem became: a beautifully constructed health graph is not particularly useful if most healthcare never enters it.
And healthcare does not arrive neatly. It happens in conversations between doctors and patients. It is written on paper prescriptions, printed by diagnostic laboratories, stored inside PDFs, passed through email attachments and WhatsApp threads, and scattered across imaging systems, hospital software and patient folders. A vital may be measured during a consultation, at home, or by a connected device.
In many parts of India, some of the richest clinical information may exist for only a few minutes as a conversation, and then disappear.
This changed how I thought about Aether. The problem was no longer simply how to organize healthcare information once we had it. The more important question became: how do we make sure clinically useful information enters the longitudinal record in the first place, without asking healthcare to first change the way it works?
Moving closer to where healthcare happens
That question pulled Aether closer to the point of care.
The patient side of Aether could reconstruct much of what had already happened. Reports and prescriptions could be uploaded, structured and added to the longitudinal history. But the patient's future was still being created somewhere else.
That led to Aether Scribe.
Ambient scribes are usually discussed as a way of reducing the documentation burden on doctors. That matters, but I became interested in something else: the clinical conversation itself is an extraordinarily rich source of information about the patient.
Symptoms, duration, medication adherence, family history, functional changes, clinical reasoning and observations may never appear in a laboratory report. Historically, much of that information has either been compressed into a short clinical note or disappeared after the consultation.
Aether Scribe turns that conversation into structured clinical information that contributes directly to the patient's longitudinal history. The point was not simply to build another note-taking tool. It was to create another way for healthcare to enter the graph.
Then reality intervenes
That immediately raises another assumption: reliable internet access.
A clinic in a smaller town or rural area should not lose the clinical record simply because connectivity is unreliable. Aether Scribe records offline and synchronizes once connectivity returns, at which point the conversation is structured and added to the patient's record.
The same principle led to something much more mundane: the scanner.
Paper is still everywhere in healthcare. Instead of waiting for clinics to stop using it, Aether Scan works around the existing workflow. A report is scanned as it normally would be. Aether picks it up, ingests it and moves the clinically relevant information into the patient's longitudinal record.
There is nothing glamorous about a scanner, and that is precisely why I find it interesting. If a technology only works after hospitals change their software, doctors change their behavior, patients change how they store records and connectivity becomes perfect, it may be technically elegant but practically irrelevant.
Healthcare already has workflows. The better question is whether intelligence can be inserted into them.
The channel is not the record
This has led to a broader way of thinking about clinical data acquisition.
A report may arrive through email. Another may be forwarded over WhatsApp. A patient may walk into a clinic holding a printed report from three years ago. A consultation may be captured through Aether Scribe. Another piece of information may already exist inside hospital software. Vitals may come from a device.
These channels matter operationally, but they should not define the patient's medical history. The channel is simply how healthcare information arrives. The longitudinal graph is where it becomes part of the patient's longitudinal history.
Instead of asking healthcare to converge onto a single input mechanism, Aether creates acquisition surfaces around the places where clinical information already exists.
Scribe is one acquisition surface. Scan is another. Patient uploads are another. Existing digital systems, messaging, email and devices are all part of the same acquisition architecture, feeding one longitudinal health graph.
What looks from the outside like a collection of features is increasingly one system: an acquisition system for clinical context. The more surfaces through which Aether captures healthcare, the denser the graph becomes. And as the graph becomes denser, the context inside it becomes richer.
What happens when enough of it accumulates
The reason this matters goes beyond creating a better medical record.
At the level of one patient, a denser graph creates richer longitudinal context. Across thousands of patients, it begins to reveal patterns that are difficult to see when healthcare remains fragmented into individual documents and encounters.
Aether already examines cohorts, co-occurring conditions, multimorbidity, patterns across clinical categories and differences between patient populations. This is population intelligence derived from the graph: a way of seeing relationships across the data that individual records make difficult to see.
The question begins to move from simply, "What has happened to this patient?" to, "What appears to be happening across these patients?"
Over time, deeper longitudinal datasets may support entirely different kinds of analysis around disease progression, treatment response, real-world evidence and eventually personalized medicine. Those are longer-term possibilities, but they begin with the same underlying requirement: enough healthcare information has to be captured, structured and connected in the first place.
Healthcare intelligence is ultimately constrained by the healthcare information available to it.
Building around healthcare as it exists
Aether now has more than 10,000 patients and more than 240,000 structured clinical parameters. It is live in two clinical deployments, with a larger hospital network in discussion.
Along the way, Tyler Cowen's Emergent Ventures at George Mason University chose to support the work with a grant.
Those are useful markers of progress. But the more important progress for me has been conceptual.
Aether began by reconstructing scattered medical records. The timeline became a graph. The graph moved closer to the clinical encounter. Conversations became something the system could capture. Poor connectivity became something the architecture had to accommodate. Paper became an acquisition surface instead of an obstacle.
Each step pointed to the same conclusion: the more healthcare Aether captures, the more useful the graph becomes.
I no longer think the difficult problem is simply organizing healthcare information after it has been created. It is building an architecture that meets that information wherever healthcare already produces it.
The health graph was only half the problem.
The other half was learning how to make healthcare visible to it.