AI Platform

The AI Platform That Powers Every THB Product

Every vendor says "AI-native." Here is what we mean: a single runtime that executes every AI interaction -- chat, voice, WhatsApp, autonomous background jobs -- through the same pipeline. The LLM proposes. A policy layer validates. Only approved actions execute. Every decision is logged. No exceptions.

0
Agent modes -- autonomous, assisted, conversational
Any LLM
Pluggable backend -- swap models without changing agents
0+
Clinical protocols grounding every response
0
Deployment modes -- chat, voice, WhatsApp, backend
Auto
Tools auto-generated from schemas -- no manual wiring
24/7
Always-on patient engagement via WhatsApp and Voice

AI Architecture: From LLM to Patient

Every AI interaction -- whether a receptionist chatting in the CRM or a background agent analyzing care gaps overnight -- flows through the same validated architecture.

LLM Providers
Foundation LLM
Primary reasoning
Voice LLM
Real-time conversations
Embeddings
Semantic search
Automatic Fallback
Multi-provider resilience
Runtime & Safety
Unified Execution
Same pipeline, every mode
Policy Engine
LLM proposes, policy decides
Trust & Safety
Injection, toxicity, topic control
Audit Trail
Every decision logged
Intelligence Layer
Clinical Knowledge
200+ protocols
Tool Registry
Auto-generated from schemas
Conversation Memory
Context across sessions
Intent Routing
Right agent, right query
Deployment Modes
Chat
CRM-embedded assistants
Voice
Real-time patient calls
WhatsApp
Async patient self-service
Autonomous
Background analysis and scheduling

Platform Capabilities

The infrastructure that makes healthcare AI safe, auditable, and production-ready -- not a chatbot bolted onto a website.

PL

Pluggable LLM Backend

Swap foundation models without changing a single agent. Multi-provider support with automatic fallback -- if the primary provider fails, the system switches seamlessly. No vendor lock-in on the AI layer.

PE

Policy Engine

The LLM never acts directly on patient data. Every proposed action passes through a deterministic policy layer that enforces permissions, validates parameters, prevents unsafe mutations, and checks rate limits before execution.

TS

Trust & Safety

Input sanitization against prompt injection attacks, toxicity filtering on every interaction, and per-agent topic boundaries that prevent scope creep. The agent handles its domain and nothing else.

KR

Clinical Knowledge Grounding

Agents reason over clinician-reviewed protocols and verified clinical data -- not training data. Hybrid search retrieves the most relevant clinical context for every query. Responses are grounded, not hallucinated.

TR

Auto-Generated Tools

Every entity in the Application Engine automatically becomes a set of tools the AI can use -- query, create, update, trigger workflows. No manual tool definitions. The AI capabilities grow as the schema evolves.

IR

Intent Routing

Inbound queries are classified and routed to the right specialist agent automatically. Goal-shift detection catches when a conversation changes topic mid-stream and re-routes accordingly.

CM

Conversation Memory

Context persists across sessions. The agent remembers what was discussed, what actions were taken, and what the patient or staff member was trying to accomplish -- even across channel switches.

CB

Cost Budgets

Per-agent cost controls prevent runaway LLM spending. Set token budgets, monitor consumption, and get alerts before agents exceed their allocation. Enterprise cost governance built into the platform.

AL

Complete Audit Trail

Every AI interaction is fully logged: what was asked, what the LLM proposed, what the policy layer approved or rejected, what executed, and what resulted. Complete traceability from prompt to patient action.

Agent Builder: No-Code Agent Creation

Hospitals can build their own agents without an engineering sprint. Define the behavior, select the capabilities, deploy to any channel.

01

Define Behavior

Plain language instructions

Write the agent's instructions in plain language. Include clinical context, safety rules, and persona. Version-controlled and auditable.

PersonaSafety RulesVersioned
02

Select Capabilities

Schema-generated tools

Choose from auto-generated tools based on your entity schemas. Test each capability independently before wiring it into the agent.

Auto-GeneratedTestable
03

Deploy & Monitor

Any channel

Deploy to any mode: chat, voice, WhatsApp, or autonomous background. Monitor via complete audit trail and action history.

ChatVoiceWhatsAppAutonomous

Four Deployment Modes, One Runtime

Same agent. Same tools. Same safety rules. Whether it's a live chat, a phone call, a WhatsApp thread, or a background job running overnight.

CH

Chat

AI embedded directly in the CRM. Agents read, write, and navigate patient records alongside the staff. Tool results rendered as interactive UI components, not just text.

Specialist agents with intelligent routing -- the right agent handles the right query.

VC

Voice

Patients call. The AI picks up. Natural conversations in Hindi and English. Books appointments, answers questions about lab results, escalates emergencies to human operators.

Real-time speech processing with natural turn-taking. No awkward pauses.

WA

WhatsApp

Patients message on WhatsApp. The AI responds with full clinical context -- appointment booking, lab results, medication refills. Safety guardrails on every interaction.

Never diagnoses. Always escalates emergencies. Messaging compliance built in.

BS

Autonomous

Agents that run without human interaction. Nightly care gap analysis, follow-up prioritization, data quality monitoring. Staff arrives in the morning to a prepared brief with recommended actions.

Scheduled and event-driven execution with reliability guarantees.

Safety at Every Layer

Healthcare AI needs predictability, auditability, and fail-safe behavior. An AI that is right 95% of the time and wrong 5% of the time is dangerous in a clinical setting.

TB

Topic Boundaries

Scope enforcement

Each agent has a defined scope. Ask it something outside that scope and it redirects or escalates. No creative improvisation on clinical topics.

RedirectEscalate
IP

Injection Protection

Input sanitization

Every input is sanitized against prompt injection attacks before reaching the LLM. Adversarial inputs are detected and blocked automatically.

SanitizedAuto-Block
OV

Output Validation

Schema-enforced

Every LLM output is validated against strict schemas before execution. Malformed or out-of-bounds responses are rejected and retried within constraints.

Schema CheckRetry
CG

Clinical Grounding

Verified protocols

All clinical responses are grounded in verified protocols and pre-computed metrics. System prompts are version-controlled and clinician-reviewed. Never hallucinated from training data.

ProtocolsVersioned

See AI in Action. Not on a Slide.

Cohort building, patient messaging, autonomous care gap analysis -- all running on real clinical data. Watch the agents work together.