Technology
Depth a chatbot wrapper can't fake.
Agastya is a production agent stack engineered to reason, speak like a human in Indian languages, and act inside your systems — on live phone calls and every text channel. This is the moat, and here is how it is built.
7
engineering layers
<1s
to first response
10+
Indian languages
1
memory across all channels
Under the hood
Seven pieces of engineering that make it work
Each one is a hard problem on its own. Together they are what let a machine hold a real conversation and actually complete the task — reliably, in-language, and cheaply enough that the economics hold.
01🧠
Agentic tool-use — it reasons, then acts
- ▸Our frontier reasoning engine runs in an agentic loop: it reads the request, decides what needs to happen, and calls your real APIs — book, look up an order, update a record, raise a ticket, send a payment link — to complete multi-step tasks, not just reply with text.
- ▸It chains steps and handles branches: check eligibility, then create the application, then confirm and message — recovering when a step returns an error rather than dead-ending.
- ▸Refusal discipline is engineered in: anything clinical, sensitive, out-of-policy or genuinely uncertain is handed to a human with full context. It would rather escalate than guess — the opposite of a bot that hallucinates a confident wrong answer.
02🗣️
Real Indian-language voice — on live phone calls
- ▸Voice runs on our native Indian-language voice engine: Indian-language speech-to-text and Indian-accent text-to-speech, tuned for how India actually speaks — natural Hindi, fluid Hinglish code-switching, and regional accents.
- ▸This is where global voice AI falls flat. Agastya understands a caller who switches between Hindi and English mid-sentence and replies in a voice that sounds local, not synthetic.
- ▸It runs on real inbound and outbound telephone calls through a carrier-grade cloud telephony webhook into the same agent brain — the voice caller and the WhatsApp user are served by identical logic.
03📚
Self-updating knowledge — multi-source RAG that never invents
- ▸A multi-source retrieval-augmented pipeline ingests your entire website plus documents — policies, price lists, FAQs, catalogues, scheme rules, SOPs — chunked and embedded into a secure, per-tenant vector knowledge base.
- ▸Answers are grounded strictly in retrieved content with page-level citations, so every response is traceable to your real information. If the knowledge base doesn't cover it, it says so and escalates — it does not make things up.
- ▸It re-crawls and re-embeds on a weekly schedule, so when your prices, policies or catalogue change, the agent relearns on its own — no manual retraining.
04🔗
One memory across every channel
- ▸A single customer identity spans phone, WhatsApp, web chat and email. A caller who followed up on WhatsApp is the same person the agent already knows — context carries across channels and across conversations.
- ▸The customer never repeats themselves and never starts from zero, whichever channel they reach you on.
- ▸That shared memory is also what feeds accurate, attributable reporting — every interaction ties back to one person and one thread.
05🔌
Deep integrations — it plugs into what you already run
- ▸Provider-agnostic adapters connect to CRM, OMS, EHR, ticketing, calendars and payment gateways — Salesforce, Zoho, HubSpot, LeadSquared, your PMS/OMS and internal APIs — behind a clean tool interface the agent calls.
- ▸It rides on your existing support number and WhatsApp Business account through carrier-grade cloud telephony — no new number for customers to learn, no rip-and-replace of your stack.
- ▸Every integration sits behind a helper that safely no-ops until it is configured, so the platform runs and degrades gracefully instead of breaking during a partial setup.
06⚖️
Human-quality at software economics
- ▸Cost-routed models: the high-volume conversation runs on a fast, low-cost model tier, while correctness-critical steps use a top-tier model — the right model for each job, not one expensive model for everything.
- ▸Prompt caching keeps each tenant's system prompt and knowledge context hot, so repeat turns are roughly an order of magnitude cheaper and faster.
- ▸Voice is metered per second and WhatsApp goes out as utility templates (far cheaper than marketing sends), so delivery cost stays a rounding error against a fee benchmarked to a human seat.
07🔒
Security & DPDP — isolation and audit by default
- ▸Per-tenant data isolation in an encrypted knowledge store with strict row-level access controls and no public policies — all access is server-side via a privileged server credential, never the browser. One client's data is never exposed to another.
- ▸Consent is captured and opt-out honoured; outbound messaging stays within approved utility templates and sensible hours, in line with India's DPDP Act. The client is the data fiduciary; Agastya processes on its behalf.
- ▸Every action the agent takes is logged as an auditable trail; secrets stay server-side and payment webhooks are signature-verified (HMAC) before any state changes. We make no security-certification claims we haven't earned.
The contrast
A chatbot wrapper vs. Agastya
Most "AI support" is a prompt wrapped around a bot on your website. Here is what the difference looks like in practice.
A chatbot wrapper
- ✕Answers with text, then dumps the user into a form or a queue
- ✕Reads one FAQ page; goes stale the day your prices change
- ✕Guesses confidently — and invents answers under pressure
- ✕Website chat only; forgets you the moment you switch channel
- ✕English-first; robotic on a Hindi or Hinglish phone call
- ✕No real hooks into your CRM, calendar or payments
Agastya
- ✓Completes the task in your systems — books, updates, pays, ticketed
- ✓Ingests your whole site + documents and re-learns weekly
- ✓Grounded in cited content; escalates instead of hallucinating
- ✓Phone, WhatsApp, chat and email with one shared memory
- ✓Natural Hindi, Hinglish and regional voice on live calls
- ✓Deep, adapter-based integrations on your existing number
The stack
The building blocks, production-grade
No mystery box — every capability is proven, and each is there for a reason.
Agentic reasoning engine
Frontier reasoning + tool-use
Native Indian-language voice
Indian-language voice STT / TTS
Secure vector knowledge base
Per-tenant vector RAG store
Carrier-grade telephony
Inbound & outbound calls
WhatsApp & messaging
Utility-template messaging
Auto-scaling secure cloud
Elastic deployment & reporting
Agentic tool-useMulti-source RAGPrompt cachingCost-routed modelsPer-tenant isolationDPDP-aware
Want the technical detail for your use case?
Get a live demo trained on your own content — we'll walk the architecture and the integration path for your specific systems.