This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
TharVA : Thar Virtual Assistant
A mobile-first, fully offline, multilingual AI assistant for camel herders in the Thar Desert.
Not a general assistant. Not a chatbot.
A field tool built specifically for Camel herders in rural Rajasthan who raise camels in one of the world's harshest environments, have no reliable internet, work with their hands, and need answers in Hindi or any other language, fast, when something goes wrong with an animal.
The spark came from time I spent in Bikaner, talking to Ashok Bishnoi, a social entrepreneur near the National Research Centre on Camel in Jorbeer and to Raika Community camel keepers whose generational knowledge of camel behavior, calving, and desert survival isn't written down anywhere accessible. What they lacked wasn't expertise. It was fast access to reliable guidance at the right moment.
One conversation stayed with me:
- A calf had been rejected by its mother.
- A time-critical emergency where the first hours determine survival.
The formal channels couldn't give clear enough answers fast enough. What actually helped was a Raika Community elder who had seen it before and knew exactly what to do.
TharVA is an attempt to make that knowledge reachable in a field, with no signal, in Hindi / any language, with one hand free.
The Two Interaction Modes
It has two interaction modes, built around how field users actually work:
Quick Call — Voice-in, voice-out. Hold a button, speak your question, hear a short direct answer. Streaming generation so TTS begins before the full response finishes. Designed for when you're standing next to a distressed animal and have thirty seconds, not three minutes.
Detailed Chat — Text or Voice input, with image support. Attach a photo of a camel's wound or a skin condition. Get a thorough, structured response. Same model, different prompt, different temperature completely different feel.
Answers are grounded in curated camel husbandry reference material from actual veterinary literature and NRCC research, injected into the system instruction at session start. The model isn't improvising from general training data. It knows the domain because it was given the domain.
How I Used Gemma 4
Model chosen: Gemma 4 E2B
Not the 26B.
Not the 31B.
The smallest one in the family — and that was entirely intentional.
The people TharVA is built for don't have high-end phones or reliable connectivity. The rule I held myself to for the entire build was: if it doesn't load and respond reliably on a mid-range Android phone in realistic conditions, nothing else matters.
The E2B — 2.3 billion effective parameters, running on as little as 4 GB of RAM, is the only model in the Gemma 4 family that makes that possible while still being genuinely capable. I have set the context length to 4096 tokens which shaped all the technical decisions I have made.
The entire inference stack runs on-device through the flutter_gemma package, wrapping Google AI Edge's LiteRT-LM runtime.
- No cloud API.
- No data leaving the phone.
- No signal required.
- For a community where privacy matters and internet is genuinely unreliable.
Offline-first wasn't a feature preference, it was the baseline.
TharVA's Application Architecture
Multimodal is no longer a premium feature
Ears (<|audio|>) — Voice input bypasses device-level speech recognition entirely. I record audio as a raw WAV file (PCM 16kHz, 16-bit, mono) and pass the bytes directly to the model. This removed the requirement to pre-install language packs through obscure settings menus that field users would never find. Unexpectedly, the E2B handled local Hindi accents and regional speech patterns from around Bikaner better than device-level ASR did. Voice input that understands your accent is voice input people will actually use.
Eyes (<|image|>) — Users can photograph a wound, a skin condition, or an animal's posture and include that in their question. I capped image support at one image per turn — a deliberate product decision, not a temporary limitation. Allowing multiple images per turn caused context overflow failures mid-conversation that were impossible to handle gracefully in the field. One image per turn gives stable, predictable behavior under real conditions.
Brain (<|think|> + system prompt) — Quick Call and Detailed Chat use the same weights but entirely different system prompts and temperatures. Quick Call prompts bias heavily toward short, direct outputs with lower sampling temperature. Detailed Chat allows longer, structured responses. The model adapts its behavior completely based on what the prompt asks. same brain, different mode.
Mouth / Vocal (streaming TTS) — I use generateChatResponseAsync() to feed tokens into text-to-speech as they arrive. The user starts hearing the response before generation finishes. Without streaming, you wait for full generation then wait for TTS. With streaming, those processes overlap. The perceived latency difference in Quick Call is the difference between an app that feels usable and one that feels broken.
Grounded knowledge (context injection) — Curated camel husbandry reference text is loaded into the system instruction at session start, truncated to a fixed character budget. Every per-turn input is kept lean, a language reminder, an optional location/battery prefix, and the actual question. The knowledge base is in context from the start without consuming fresh tokens on every turn. This was forced by the 4,096-token on-device context limit, which is the real constraint that shaped almost every other technical decision.
Multilingual behavior — Any language or mode change triggers a full session reset: close the inference session, rebuild the system prompt with fresh language reminders, start fresh. Without hard resets, KV cache state bleeds between contexts, the wrong script, the wrong tone, the wrong response length in ways that undermine trust in the app entirely.
The hardest engineering in this project was completely unglamorous: a download recovery system that detects partial model files and restarts cleanly, a runtime compatibility fix for a silent mismatch between the LiteRT-LM version and the updated Hugging Face artifact format, and a turn cap that forces session rotation before context overflow causes silent failures mid-conversation.
None of that shows up in a demo. All of it determines whether the app actually works in a field in Rajasthan.
Code
https://github.com/daathwi/TharVA
The Raika community have kept camels alive in the Thar Desert for centuries. They don't need an AI to tell them what they already know. What TharVA tries to do is make the knowledge that exists in community memory and veterinary literature reachable at the moment when someone needs it with no signal, in Hindi, with one hand free.
That's a narrow goal. I think it's the right one.