A plain-language guide for the clinical team · July 2026
Our platform isn't one chatbot — it's a small team of specialized AIs, each with a
defined role, defined limits, and a defined way of making decisions. Like colleagues,
some talk to patients, some watch quietly in the background, and one is the strict
clinical brain the others consult.
Everything on this page describes what is live today, except where
marked otherwise.
Working with patients
The Health Companion
Live
The chat (and voice) assistant patients talk to daily. Think of it as a
well-briefed health coach who has actually read the chart: before
answering, it pulls that patient's own CGM, meals, activity, sleep, and documents —
it never answers from general assumptions.
When it acts
Whenever the patient asks — in the app, by voice, or on WhatsApp.
Voice understands major Indian languages, not just English.
What it remembers
Facts patients share ("I'm vegetarian", goals, allergies). What a
patient explicitly states always outranks anything the AI merely inferred.
What it won't do
Prescribe, adjust doses, or endorse stopping medication — always
redirected to the care team. An active hypo gets rule-of-15 first aid
immediately, never a brush-off.
What to expect
Answers framed to that patient's condition and goals —
glycemic language for CGM patients, weight/satiety framing for weight-loss
patients. Missing data is stated, never invented.
How it decides — every single question
1Understand — what is being asked,
about which data, over which dates. Urgent symptoms skip straight to a full answer.
2Investigate — pulls only this
patient's records; harder questions get a deeper multi-specialist review
(glucose, nutrition, fitness, sleep each examined separately).
3Check the evidence — every claim
must trace back to a record it actually retrieved; gaps are declared.
4Respond — warm, specific, and
honest about uncertainty. Correlations are hedged ("lines up with"), never
declared as proof.
The Watchful Eye (proactive monitor)
Live
The one patients don't talk to — it watches and reaches out first.
It's behind every notification a patient receives: the morning daily brief, a nudge
after a logged meal, an alert after an overnight low.
When it acts
Scheduled check-ins up to 4× a day, only between
7 AM–10 PM in the patient's own timezone — plus instant reactions
(within ~30 seconds) when a meal, glucose reading, or symptom is logged.
How it decides severity
A pattern seen once is info. Persisting 3 days → attention;
5 days → warning. Real-time glucose threshold crossings are treated as
safety-relevant immediately.
Anti-spam by design
Never repeats the same topic within 24 hours, respects a
daily notification budget, and sends only the single most important
insight per scan.
What it won't do
Raise alarms on a healthy day — false alarms teach patients to
ignore notifications, so "all good" days get encouragement or silence.
Domains without data are never mentioned.
The Meal Analyst
Live
When a patient snaps a photo or describes a meal, this agent identifies the foods,
estimates portions and nutrition, and returns a 0–100 score where every
deducted point has a written, cited reason — visible in the app.
How it decides
Concerns and positives must cite a source: the meal's own numbers,
the patient's plan, their history, or a named guideline. Uncited
commentary is discarded before the patient sees it.
Condition-aware scoring
Glycemic load carries full clinical weight only for glycemic
patients (reviewed & signed off, July 2026). A weight-loss patient's dessert
is scored on calories and satiety, not clinical glucose risk.
Glucose prediction
A predicted rise is shown only when the patient has real CGM
history to predict from. No history — no number, by rule.
Honest uncertainty
"Some snacks at the office party" is logged as a low-confidence
estimate — it will not invent a confident itemized menu.
The clinical brain they all consult
The Metabolic Engine
Live
Not a chatbot at all — this is deterministic clinical logic plus a validated
prediction model. The same inputs always produce the same outputs, which is
why the conversational agents consult it rather than reason about metabolism themselves.
What it judges
Whether a glucose pattern is meal-driven or
physiology-driven — so patients aren't blamed for spikes their meal
didn't cause. Also body-composition trends from InBody data.
Permission to advise is earned
It may make a suggestion only when the pattern is
meal-driven, there's enough history behind it, and it can cite a specific lever.
Otherwise it is restricted to stating facts — or stays educational.
It audits itself
Every piece of advice is entered in a ledger; days later it checks
the patient's actual CGM response against what it predicted — a running record of
whether its advice works.
Safety override
Any safety signal pre-empts all coaching. Safety messages
always outrank optimization tips.
Working for you, the clinical team
The Panel Assistant
Live
The provider-side of the Health Companion: ask about your assigned
patients in plain language — "prepare a summary for my appointment with
Asha", "how has she been sleeping?" — and it reviews their full record the same
way, with clinical rather than coaching framing.
Access is enforced
You can only reach patients assigned to you — checked on
every request, not on trust.
Cohort questions
A research view can answer across a panel — "which of my patients
had hypos this week?", GMI-based classification of who's trending toward
diabetic / prediabetic ranges.
The Dashboard Guide
Live
A small helper that answers "where do I find…?" questions about the provider
dashboard — reports, packages, patient history tabs — so nobody needs a manual.
Describe Your Day (voice day-logging)
In design
The next addition: a patient says out loud how their day went — "woke at 7, had poha
around 9, walked after lunch, headache in the evening" — and the AI asks clarifying
questions, then proposes the entries on their timeline. Nothing is saved
until the patient confirms each item. The groundwork (correct handling of
after-the-fact logging) is already built and tested.
What holds it all together
One shared memory, with rules. All agents read the same patient
facts, and the same law applies everywhere: what a patient explicitly said can never
be overwritten by an inference. Records separate when something happened
from when it was recorded — like proper charting.
Nightly examination. Every agent on this page is tested nightly
against a bank of standardized scenarios — including deliberately dangerous ones —
with a zero-tolerance rule on safety cases. A failure automatically files an
incident for follow-up.
Your feedback becomes the test. Responses you flag are reviewed
monthly; genuine problems become new scenarios in the test bank, so the same
mistake cannot recur.