AI architecture
Pulse OS is built so Pulsey cannot invent a reading. Retrieval, grounding, and access control run before a reply ships. Text in the product today.
Core architecture
Notes, document chunks, and conversational memory sit in Postgres with pgvector. Similar context is retrieved with an HNSW index before Pulsey answers. The agent reasons over what was fetched, not over a remembered typical value.
Every number in a Pulsey reply has to appear in a tool result or in the person's own words. If it does not, the reply is regenerated once and then replaced. That is what makes the product's claim auditable.
A coaching query cannot bind labs. Clinical notes sit on the shared session. The athlete grants each role. Enforcement is at fetch time, before anything reaches the model.
A value from a PDF stays in review until a clinician approves it. Unapproved biomarkers do not feed a finding, a flag, or a coaching answer.
Missing wearable days stay empty. Nothing is interpolated. If a device did not write a point, Pulsey says there is no data yet.
On the desks
Open Talk to Pulsey on each desk. A coach asking for ferritin is refused. A clinician gets the visit brief, with the lab still in review. An athlete can log how something felt.
Pulse OS
HRV is below her own trailing range. Cycle is context. Labs are not in this query.
Pulsey overnight list
Amara below her own HRV median for a third day. Priya at 1.6x usual load. Nadia unsynced four days. Ask Pulsey for the list. Bloodwork stays out.
HRV vs her 28-day median
48ms
Session load this week
36
Her week · you wrote this
Open a session for load and the recovery that followed.
