The agent, on a phone,
in every language.
Tshomi speaks first, in the participant’s language. Every message calls a tool. Every tool leaves an audit-trail row. 98% of sessions close without a human.
Not a chatbot.
A tool-calling agent.
Tshomi doesn’t just reply. Every turn either calls a real tool (book a slot, verify a document, record a check-in, deliver a payslip) or hands off to a supervisor with the context to act. Nothing it does is invisible.
WhatsApp for participants.
In-app for supervisors.
Same data, same audit trail. Different surface for each audience.
Participants never install anything. Tshomi lives in the WhatsApp thread they already trust — same phone number they use for family and friends.
Supervisors get a full operations console — same data, same audit trail, richer view. Runs as a PWA in the browser or a native app on their device.
Every message calls a tool.
Every tool call is a row in the audit log — exportable JSON, filterable by programme, cohort, participant or period.
What Tshomi does, without a human.
Every participant is met in the language they set — from the first message. The language switches; the tool call and receipt don't.
Check-ins, payouts and consent-gated flows require a biometric match. Templates encrypted at rest, revocable on request.
Bank letters, ID copies, contracts — Tshomi asks for them, receives them in WhatsApp, and verifies them before writing the audit row.
All nine SA official.
Plus English and Afrikaans.
Participants are greeted in their mother tongue on the first message. Same tool call, same audit-trail row — the language switches, the receipt doesn’t.
Model stack.
Guardrails.
Haiku for triage. Sonnet for narrative.
Fast, cheap model handles routing, classification and structured extraction. Larger model runs for participant-facing narrative in-language. Guardrails at both layers.
Randomised chat data. Zero PII in logs.
- PII redacted before any model call
- Structured outputs validated against JSON schema
- Tools gated by programme-scoped RLS
- All prompts version-controlled and diffable
Book a live demo.
30 minutes. We walk your team through a live thread, show the tool-call log in real time, and generate a compliance pack from your own sample data.