Sarvam AI's fourth Build with Sarvam roundup, posted August 19, 2026, skips the usual product-launch format and just shows what builders are already shipping on its stack. Three projects made the cut: a Raspberry Pi device that fakes a full household to scare off burglars, a voice agent that haggles over vegetable prices in three languages at once, and a CCTV copilot that searches surveillance footage by conversation instead of by scrubbing through hours of recording.
If you've been tracking Sarvam's public launch of Voice Agents or its Bulbul v4 text-to-speech model, these three builds are a useful answer to the question every API launch leaves open: what does this actually look like once real builders get their hands on it?
TL;DR
| Build | Builder | What it does | Built on |
|---|---|---|---|
| Fake Family | Pankaj (@the2ndfloorguy) | Raspberry Pi device that simulates an occupied household to deter burglars | Sarvam Text-to-Speech |
| Hyderabadi Sabzi Aunty | Vishal Singh (@vishalsingh2972) | Voice agent that bargains for vegetables in mixed Hindi/Telugu/English | Sarvam Voice Agents |
| Drishti | Nithin Varghese (@call_me_Nithin) | Regional-language CCTV investigation copilot | Sarvam stack, built at Sarvam Epoch Buildathon |
Fake Family: a burglar deterrent built from a TTS API
Pankaj built Fake Family after thieves started targeting empty houses in his hometown. Instead of a traditional alarm, the device — running on a Raspberry Pi — uses Sarvam's Text-to-Speech model to generate the ambient sound of a lived-in Indian household: husband-and-wife conversation and arguments, TV audio, and cooking and chai-making sounds, layered with household lights switching on and off on a real schedule.
What makes this a genuinely hard TTS problem, not just a novelty: generic single-speaker, single-sentence speech synthesis doesn't get you here. Convincing a passerby that a house is occupied means generating ambient, multi-voice, contextual audio that sounds like real domestic life rather than a voice assistant reading a script — a much rougher test of a TTS model's naturalness than a typical demo.
Hyderabadi Sabzi Aunty: code-switching under pressure
Vishal Singh built Hyderabadi Sabzi Aunty the same week Sarvam opened public access to Voice Agents — the product that had already logged over 350 million conversations in earlier enterprise deployments before its public builder launch. His agent plays the part of a vegetable vendor a caller can haggle with, in a live mix of Hindi, Telugu, and English.
The technically interesting part isn't the roleplay — it's what the agent has to handle to make that roleplay feel real:
- Interruptions mid-response, without losing conversational state
- Mid-sentence language switching, the way real code-switching actually happens in everyday Indian speech, not clean per-turn language selection
- Persistent memory of the order across a multi-turn negotiation
- Real-time total calculation as items and quantities change
This is a harder bar than most Western voice-agent demos clear, because code-switching mid-sentence — not just alternating languages between turns — is normal in everyday Indian conversation and genuinely difficult for a model trained primarily on monolingual turns to track without breaking context.
Drishti: making CCTV footage searchable in your own language
Nithin Varghese built Drishti at the Sarvam Epoch Buildathon — the builder-track finale where ten teams presented working products on the Sarvam stack. Drishti is a regional-language investigation copilot for CCTV footage: instead of manually scrubbing through hours of recording, an investigator can ask questions in Hindi, Malayalam, or another regional language and have the system search video conversationally.
This points at a use case well outside the usual voice-agent demo set: video content that becomes queryable through natural-language, regional-language questions, aimed at law enforcement and security teams working in a language other than English. It's the kind of application that a general-purpose, English-first voice stack tends to handle worst — which is exactly the gap sovereign, India-first models are built to close.
Why these three builds matter for voice-AI builders generally
None of these three are hackathon toys that stop working outside a demo. Each one stress-tests a capability that's genuinely difficult to get right in production voice AI:
- Ambient, multi-speaker audio generation (Fake Family) — beyond clean single-voice TTS.
- Mid-sentence code-switching with persistent state (Sabzi Aunty) — beyond turn-based language selection.
- Conversational, regional-language search over non-text media (Drishti) — beyond English-first retrieval.
If you're building voice agents for any market where users don't speak a single language cleanly by default — which is most of the world outside a narrow English-first assumption — these are the exact failure modes worth testing for before you ship, not after.
Related on explainx.ai
- Sarvam Voice Agents Are Public: What Builders Can Actually Use — the product these builds run on, and what changed with the public launch.
- Sarvam AI: Full Capabilities Guide — Models, API, Speech, Vision — how to get started with Sarvam's full API stack.
- Bulbul v4: Sarvam's More Expressive Indian Voice Model — the TTS model behind Fake Family's ambient audio.
- Sarvam Epoch 2026: Every Confirmed Launch — the buildathon where Drishti was built.
- India's Sovereign AI Status: IndiaAI Mission — the broader policy push behind India-first AI stacks like Sarvam's.
- Build a Speech-to-Speech Voice Agent — a hands-on guide for building your own voice agent.
- Types of AI Agents: Complete Guide — where voice agents fit among broader agent categories.
Source: @SarvamAI on X, Build with Sarvam edition 4, August 19, 2026.
This post reflects the builds and details Sarvam AI shared publicly as of August 19, 2026. Availability, pricing, and product details for Sarvam's underlying APIs may change — verify current specifics on Sarvam's official site before building on them.