Built on Gnani Prisma and Timbre · 10 Indian languages
CareLoop calls patients after they leave, runs a condition-specific check-in, triages what it hears in real time, books the follow-up visit during the call, and hands your care team a structured report.
Same call, 10 languages
Starts in English, then walks through every language. Tap one to jump to it.
CareLoopAarogya Demo HospitalAarogya Demo Hospital, Whitefield
Hi Suresh, a 2-minute check-in on how you are recovering.
Automated assistant, not for emergencies. Call 108 · Synthetic demo
What the care team gets
Whatever the language, answers land as the same typed fields, ready for rules, reports and FHIR export.
Speaks the language your patients think in
Demo
A two-minute walkthrough of the product in action, using synthetic data.
Synthetic data only. Loads when you press play.
How it works
One loop: call, understand, escalate, schedule. Your team only sees the patients who need them.
Day 1 and day 7 after surgery, day 7 for dermatology, day 30 for chronic care. Questions come from condition packs, so you get typed answers like itching 0-10 or medicine adherence, not a vague transcript.
Opens with a consent disclosure in the patient's language.
Red flags are caught in the moment. On an urgent or emergency flag the agent speaks a scripted safety line, a nurse alert is pushed live to the dashboard, and the exact words that triggered it are recorded.
Triage can only raise urgency, never lower it.
Book, reschedule or cancel a follow-up against real slots, route to another department when the complaint doesn't match the specialty, and track per-medicine adherence with a doctor and pharmacy flag after repeated misses.
No answer? Retry, then a WhatsApp or SMS link to the same check-in.
Priority
Alerts stream in live from calls in progress and sort by urgency, not arrival time. A nurse acknowledges or assigns with one click.
Open alerts
0
Urgent or emergency
0
Flagged this week
12
Illustration with synthetic data. Every alert carries the rule that fired and the patient's own words.
Safety
Three independent layers watch every call. Any one of them can escalate. None of them can talk another one down.
no fever, but I have chest pain
Every final transcript
All 10 languages plus romanised Hinglish. Clause-aware negation, so "no fever" is not a flag. A negated emergency phrase is downgraded to REVIEW, never dropped.
Waiting…
After each typed answer
Deterministic expressions such as swelling and fever → URGENT, or pain score ≥ 8 → URGENT.
Waiting…
End of call
A clinician summary plus a second opinion that can raise the level, never lower it.
Waiting…
Scheduling
The agent offers real slots from your doctors' calendars and books, reschedules or cancels before the call ends. Two calls reaching for the same slot can't both get it.
9:30 am
Booked
10:00 am
Free
10:30 am
Free
11:00 am
Booked
11:30 am
Free
2:00 pm
Free
2:30 pm
Booked
3:00 pm
Booked
3:30 pm
Free
4:30 pm
Free
On the line right now
Waiting for the next call…
Reports and transcripts
Every call becomes typed answers, the reasons it was flagged, what was done, and a transcript that points straight at the turn that mattered.
Abhishek Ghosh is post-appendectomy and reports a pain score of 2/10, which is improving. He reports a slight fever but notes the wound appears clean and dry. He is mobile, denies nausea or vomiting, and is adherent to all medications.
Why this was flagged
REVIEWFever mentionedlive red-flag scan
slight fever but wasn't that much
REVIEWPost-operative feverpack rule surg_fever
slight fever but wasn't that much
Answers
On a scale of 0 to 10, how much pain do you have right now?
2
“two”
Is the pain getting better, staying the same, or getting worse?
Better
“it is getting better”
Have you had any fever or chills?
Yes
“slight fever but wasn't that much”
How does the wound look?
Clean dry
“it is clean and dry”
Are you able to walk around normally?
Normal
“i am able to walk around normally”
Are you taking all your prescribed medicines?
All
“all all”
Recovery
0.0/ 100
Actions taken
No answer
CareLoop retries on a schedule, then sends a WhatsApp and SMS link that opens the same check-in, by voice or text, in the patient's language.
Attempt 1
Ringing…
Attempt 2
Retries 2 hours later
Attempt 3
Retries 2 hours later
After the last attempt
Up to 3 call attempts first.
Condition packs
Record a discharge and the right pack schedules the calls. Each pack asks typed questions and carries its own red-flag rules, so the output is data, not a vague transcript.
Calls on day 1 and day 7 8 questions 9 red-flag rules
It asks
Escalates when
Analytics
Answer rates, completion, medicine adherence, response latency and recovery trends by condition, so you can see where patients need more support.
Calls placed
0
Answer rate
0%
Completion rate
0%
of answered calls
Medicine adherence
0%
doses fully taken
Median latency
0 ms
voice calls
Illustration with synthetic data.
And more
The conversation is one part. The rest of the loop is built in too.
Hinglish, Tanglish and Benglish go through Prisma v3.0 with auto-detect and keyword hints for doctor and medicine names.
One Timbre voice per language. Patients can interrupt, and their mic is muted while the bot speaks so echo isn't answered.
Every check-in can be taken as a call or typed in chat, from the same link.
Per-medicine tracking, with a doctor and pharmacy flag after repeated misses.
When the complaint doesn't match the original specialty, the agent routes the patient to the right department.
Record a discharge and the matching pack schedules the follow-up calls automatically.
QuestionnaireResponse and Observation export for HIS and EMR import.
Every report view and export is recorded and visible to admins.
Nurse, doctor and admin views. Consent is recorded before any answer is stored, and can be withdrawn.
Privacy and compliance
All data in this demo is synthetic. No real patients, phone numbers, IDs or recordings are used.
The call opens with a disclosure in the patient's language. No health answer is stored until consent is recorded.
Role-based access for nurses, doctors and admins. Audio isn't stored by default.
Every report view or export is written to the audit log, visible to admins.
Patients can withdraw consent or request erasure, designed around the DPDP Act 2023.
FAQ
English, Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Tamil and Telugu, the 10 languages Gnani Prisma and Timbre support. Code-mixed speech such as Hinglish, Tanglish and Benglish is handled by Prisma v3.0.
Triage can only raise urgency, never lower it. Every transcript is scanned by a clause-aware red-flag lexicon in all 10 languages, deterministic pack rules run after each answer, and an end-of-call LLM review can raise the level but not reduce it. On an urgent or emergency flag, a scripted safety line is spoken verbatim, with no LLM improvisation, and a nurse alert is raised.
CareLoop retries on a schedule, then falls back to a WhatsApp or SMS link that opens the same check-in by voice or text, so a missed call doesn't become a missed follow-up.
The call opens with a consent disclosure in the patient's language, and no health answer is stored until consent is recorded. Access is limited to care-team roles, every report view or export is audit-logged, and patients can withdraw consent or request erasure. CareLoop is designed around the DPDP Act 2023.
Reports export as FHIR R4 (QuestionnaireResponse and Observation) and as CSV, so they can be imported into HIS and EMR systems.
No. This is a hackathon build for the Gnani AI Great Indian AI Internship Challenge, and all data in the demo is synthetic.
Open the demo dashboard, pick a patient, and take the call yourself in text or voice. Everything in it is synthetic.
Open the live demo