B-CON AI
AI that hears, speaks, reasons and sees. In Nepali. On our servers.
सुन्ने, बोल्ने, सोच्ने र देख्ने एआई। नेपालीमा। हाम्रै सर्भरमा।
Four models we train and serve ourselves, built for the way care actually happens in Nepal. No patient data goes to a third-party AI API. Every product in the family uses them.
- models in production
- 4
- Nepali speech to text
- ~1 s
- third-party AI APIs
- 0
- reviewed and signed by a clinician
- 100%
01The models
Four senses, one patient.
Each model does one thing well and hands off to the next, the way a clinical team does.
- 01 · hearing
B-CON Ears
Nepali speech recognition
Hears Nepali the way it is spoken in a clinic.
Speech to textlive - 02 · speaking
B-CON Voice
Nepali speech synthesis
Speaks Nepali like a person, not a robot.
Text to speechlive - 03 · reasoning
B-CON Brain
Clinical language model
Writes the note, drafts the plan, answers the question.
Language and reasoninglive - 04 · seeing
B-CON Eyes
Clinical document vision
Reads prescriptions, reports and cards from a phone photo.
Vision and document understandinglive
02In the consultation room
From spoken Nepali to a signed note.
This is the flow running in clinics today, not a concept.
- डाक्टर · B-CON Earsनमस्ते, के भयो?Hello, what brings you in?
- बिरामीतीन दिनदेखि ज्वरो आएको छ, टाउको दुख्छ, जीउ पनि दुख्छ।Fever for three days, headache, body aches.
- डाक्टर · B-CON Earsबान्ता भयो? रगत बगेको छ कतै?Any vomiting? Any bleeding?
- बिरामीबान्ता छैन, तर खाना मन लाग्दैन।No vomiting, but no appetite.
- डाक्टर · B-CON EarsDengue को लागि CBC र NS1 गरौं। Paracetamol 500 mg 1-1-1, पानी धेरै पिउनुहोस्।We will do a CBC and NS1 for dengue. Paracetamol 500 mg 1-1-1, drink plenty of water.
- Chief complaint
- Fever x 3 days with headache and myalgia
- History
- No vomiting, no bleeding, reduced appetite
- Assessment
- Acute febrile illness; dengue to be excluded (ICD-11 1D2Z)
- Investigations
- CBC with platelets · Dengue NS1 antigen
- Plan
- Paracetamol 500 mg 1-1-1 x 3 days · oral fluids · review in 48 h or earlier if bleeding
Rulebook
- ✓Drug interactionsnone
- ✓Allergiesnone recorded
- ✓NSAIDsavoided in suspected dengue
- ✓Dose by weightwithin limit
Signed · Dr. Sharma · 14:32
Illustrative consultation. Names are fictitious; the flow is the product.
03How it is hosted
Inside the fence.
Products talk to one B-CON AI gateway. The gateway removes identifiers, routes to the right model on our GPUs, enforces the output schema and the safety rulebook, and returns structured data. Audio and images are processed in memory and discarded.
- Products
- B-CON EHR
- B-CON EHMIS
- B-CON PHR
- B-CON Pharmacy
- B-CON Blood Bank
- B-CON Disease Surveillance System
- B-CON Maternal and Child Health Management System
- B-CON NCD Screening and Monitoring System
- School Health Information System
- Provincial Health Profile
- CGPP Nepal Connect
- Kansuke
B-CON AI gateway
- 1Remove identifiers
- 2Route to model
- 3Enforce output schema
- 4Run the rulebook
- 5Audit the call
Audio and images are processed in memory and discarded. Nothing is used for training.
- Models on our GPUs
- B-CON EarsSpeech to text
- B-CON VoiceText to speech
- B-CON BrainLanguage and reasoning
- B-CON EyesVision and document understanding
04Principles
Rules we do not bend.
- 01
The clinician decides
Every note, plan and suggestion is a draft until a doctor reads it and signs it.
- 02
Interactions from a database
Drug-drug interactions, allergies and dose limits come from a curated rulebook. The model explains; it never invents.
- 03
Structured, not chatty
Outputs are schema-enforced JSON that lands in the right field, orderable from the hospital's own catalogue.
- 04
Nepali first
Trained and evaluated on Nepali speech and clinical language, with English terms where clinicians use them.
- 05
Nothing leaves
Self-hosted inference, identifiers removed, no data used to train anything.
- 06
Measured, not promised
Every model has an evaluation set and a latency budget, and we publish what we measure to customers.
FAQ
Questions clinicians ask
Will it prescribe for me?
No. It drafts. Suggestions are matched to your catalogue and checked against the rulebook, and nothing is ordered until you sign.
What happens to the audio?
It is transcribed in memory on our servers and discarded. The transcript is part of the encounter you control, and identifiers are removed before the language model sees it.
Does it work in our dialect?
B-CON Ears is trained on Nepali from across the country and improves on each hospital's vocabulary with their permission. We measure accuracy with you before switching it on.
Can we use the models outside B-CON products?
Yes, through an API agreement for Nepali healthcare organisations. Ask us.
Hear it in Nepali.
A 30-minute call is enough to show a consultation end to end.