Pre-launch draft. Demo encounters are synthetic and have not yet been reviewed by a clinician; accuracy figures are not yet published.

Built for code-switched clinical speech

Your patients don't
switch to English.
Neither should your notes.

Bolo listens to code-switched Indic-English clinical speech — the kind every model in this market was trained to ignore — and produces a draft note, a routed call, or a follow-up in the patient's own language.

Endocrinology follow-up · 58F · Gujarati + English

Not clinically reviewed

No medication change is ever written to the chart by the system. Draft only; the clinician signs.

Synthetic demo encounter. Not a real patient.

Book a pilot call See how Scribe works

The gap

One utterance, two very different results.

What the patient said

મને છાતીમાં chest pain થાય છે since બે દિવસથી.

Mane chhātīmā chest pain thāy chhe since be divasthī.

English-first model

chest pain since

Duration lost. Onset lost. The clinical content of the sentence was in Gujarati.

Bolo

58F reports chest pain × 2 days.

Grammar in Gujarati, clinical noun in English, one structured result.

An English-first model is not broken here. It is doing exactly what it was built to do: find the English and discard the rest. The result is a note that is confidently incomplete, which is worse than one that is obviously wrong — nobody reviews a sentence that reads fine.

Three products, one platform

Three pains. One wedge, three times.

Practices buy these one at a time, and they should. The common technical problem underneath all three is the same: real Indic-English conversation switches language inside a clause, and every incumbent is English-first.

The problem, in the practice's words

Your physicians finish clinic and start documentation. On a panel that speaks Gujarati, Telugu or Hindi at home, they are also doing a second job during the visit: translating in their head while they listen. The ambient scribes on the market help with the first problem and make the second one worse, because they transcribe the English and lose the rest.

P2

Bolo Scribe

Ambient documentation of a code-mixed encounter.

Risk class · Documentation

How Scribe works

The problem, in the practice's words

The front desk is the first place the language gap costs the practice money. Calls take longer, get handed between staff to find someone who speaks Gujarati, and are abandoned when nobody does. The patients who hang up are the ones least able to get care another way.

P1

Bolo Front Desk

A receptionist that answers in the caller's language.

Risk class · Administrative

How Front Desk works

The problem, in the practice's words

Post-discharge call volume lands on the nurses, and the calls that most need making are the ones hardest to make — a post-op patient who speaks Hindi, at home, on day three, with a question about a dressing. Those calls get shortened or skipped, and the readmission is the consequence.

P3

Bolo Follow-Up

Post-visit check-ins that answer from your own instructions.

Risk class · Clinically adjacent — highest risk

How Follow-Up works

How the speech layer works

Four steps, and the second one is the whole company.

  1. 01

    Frame-level language ID

    Every 40 ms, as a conditioning signal into the decoder — not as a routing decision. A router has to commit before the switch happens, and is wrong by construction mid-clause.

  2. 02

    Code-switch decode

    One multilingual acoustic model with a shared token inventory, not several monolingual models behind a switch. The hard case is an entity that straddles the switch point — “blood pressure થોડું high છે” — where a router returns half a fact.

  3. 03

    Entity linking and confidence

    SNOMED CT, RxNorm, ICD-10, LOINC. Every quantity, dose, frequency and vital is scored separately from the text around it.

  4. 04

    Structured output

    A schema the model fills, never free prose into a chart. Every assertion carries a provenance pointer back to the audio.

What we publish

We have not published accuracy numbers yet.

Language coverage and methodology

The eval harness is built and the numbers are not ready. Publishing a figure we cannot stratify by language and code-mix band would be worse than publishing nothing, because the aggregate number is the one that hides the failure. When these tables have real values in them, they will name the eval set and the date.

If you are evaluating us, ask for the current run. We will send it whether or not it flatters us.

Safety posture

Deterministic core, generative edges.

Anything that writes to a chart or books a slot runs through a state machine with explicit confirmation. The language model writes prose; it never writes facts it invented. This is not conservatism — it is the only architecture that survives a deposition.

Draft only, always
The clinician signs. There is no configuration flag that changes this.
No auto-committed medication changes
Low-confidence quantities render amber and must be touched by a human.
Nothing untraceable ships
Every clinical assertion carries a timestamp and a confidence score back to the transcript span that produced it.
Escalation is the default
In Front Desk and Follow-Up, anything clinical goes to a person — tuned for recall, not precision.

Security

The questions your security review will ask, answered on one page.

Including the two we have not settled yet, which are named on the page rather than left out.

Read the security page

Pilot

Book a pilot call.

Thirty minutes. We will run the demo on the languages your panel actually speaks, show you the eval methodology, and tell you plainly which parts are not ready. If we are not a fit we would rather find out on this call.

One reply from a person. We do not run a nurture sequence and we do not share your address.