krishivaani-gemma2-2b

KrishiVaani — కృషి వాణి

A safety-grounded voice AI for Telugu-speaking farmers.
A farmer speaks a question in rural Telugu. They get spoken advice back in under 3 seconds. No smartphone required. No literacy required. No internet beyond 2G.

---

The Problem

Telangana has 5.8 million smallholder farmer households. When a crop gets sick, the nearest expert is the pesticide shop owner — whose advice is commercially motivated and often dangerous. Farmers are routinely told to use banned chemicals, apply overdoses, or spray days before harvest. This is a primary driver of farmer debt in India.

Existing agricultural AI is built for smartphones and English. It doesn't reach the people who need it.

---

What KrishiVaani Does

Voice-in, voice-out advisory pipeline in Telugu. The key technical contribution is a post-LLM safety verification layer that intercepts every response containing chemical names or dosages, cross-checks them against an ICAR ground-truth database, and corrects or blocks them before the farmer hears the answer. It is a hallucination kill-switch for chemical advice.

┌─────────────────────────────────────────────────────────────────────┐
│                        KrishiVaani Pipeline                         │
└─────────────────────────────────────────────────────────────────────┘

 WAV upload (2G)
      │
      ▼
 ┌─────────┐   low confidence   ┌───────────────┐
 │   ASR   │ ─────────────────► │ Clarification │──► TTS ──► WAV out
 │ Whisper │                    └───────────────┘
 └─────────┘
      │ ok
      ▼
 ┌─────────────┐
 │ Memory Load │  (SQLite, last 6 turns)
 └─────────────┘
      │
      ▼
 ┌────────┐  OUT_OF_DOMAIN ──────────────────────────────────► TTS
 │ Router │  WEATHER / MARKET ──► ┌───────────┐
 │        │                       │ MCP Tools │──► OpenWeatherMap
 └────────┘                       │           │──► Agmarknet API
      │ DISEASE / FERTILIZER /    └───────────┘
      │ IRRIGATION / GENERAL           │
      │◄───────────────────────────────┘
      ▼
 ┌──────────────────┐
 │  LLM             │  Gemma-2-2B (QLoRA fine-tune, served via HF/Groq)
 │  + local Telugu  │  → deterministic fallback when model unavailable
 │    advisory      │
 └──────────────────┘
      │
      ▼
 ┌──────────────────────────────────────────────────────┐
 │  ICAR Verification  ← the novel part                 │
 │                                                      │
 │  • banned chemical in response?  → block, replace    │
 │  • mentioned dose > ICAR max?    → correct in-place  │
 │  • pesticide + near-harvest?     → append PHI warning│
 │                                                      │
 │  runs only on DISEASE + FERTILIZER (no false alarms) │
 └──────────────────────────────────────────────────────┘
      │
      ▼
 ┌─────────────┐
 │ Memory Save │
 └─────────────┘
      │
      ▼
 ┌─────┐   Edge TTS (te-IN-ShrutiNeural)  → 8kHz WAV ──► farmer
 │ TTS │   gTTS fallback
 │     │   pyttsx3 offline fallback
 └─────┘

---

Evaluation

Safety (ICAR Verification Layer)

Test set: 30 hand-labelled cases — 10 banned chemicals, 8 overdose responses, 3 near-harvest, 9 safe queries.

The ICAR layer is the only component that matters for safety. Ablation: remove it → all metrics go to 0%.

Latency

Measured end-to-end on text pipeline (LLM + routing + memory, no external TTS):

Accuracy (ROUGE-L)

Mean ROUGE-L: 0.071 against a 35-item held-out Telugu agri Q&A set.
This is a baseline. The fine-tuned Gemma-2-2B is not yet uploaded — the system currently runs the deterministic Telugu advisory fallback. Fine-tuning on more KCC data is the next step.

---

Quickstart

git clone https://github.com/asheesh07/krishivaani.git
cd krishivaani
pip install -r requirements.txt
cp .env.example .env
# fill in HF_TOKEN at minimum
uvicorn main:app --reload

Check what's configured:

GET /ready

Send a text query (no audio needed to test):

curl -X POST http://localhost:8000/chat/text \
  -H "Content-Type: application/json" \
  -d '{"user_input": "వరికి యూరియా ఎంత వేయాలి?"}'

Send a voice query:

curl -X POST http://localhost:8000/chat/voice \
  -F "[email protected]" \
  --output response.wav

---

API

All voice output is 8kHz mono WAV — tuned for 2G bandwidth and basic Android media players.

---

Architecture Decisions

Why LangGraph instead of a monolithic prompt?
Each node has a single responsibility and can fail independently. The ICAR verification node intercepts after the LLM regardless of what the LLM decided — you can't do that cleanly inside a single prompt chain.

Why a rule-based router instead of LLM-based intent classification?
Latency and reliability. The keyword classifier adds <1ms and has zero failure modes. For a 7-class problem with Telugu agricultural vocabulary, keyword matching is sufficient and auditable.

Why SQLite instead of a vector store?
Farmers ask sequential follow-up questions in a session, not semantic searches across history. Ordered turn storage with a 6-turn window is the right model.

Why 8kHz output audio?
Nyquist limit for voice intelligibility is ~4kHz. 8kHz WAV is half the size of 16kHz with no perceptible quality loss for speech on phone speakers.

Why does the ICAR layer only run on DISEASE and FERTILIZER intents?
Running it on irrigation or weather queries would produce false positives (e.g., flagging "nitrogen" in a weather context). Scope restriction is how you get 0% false positive rate.

---

Configuration

Copy .env.example to .env. Only HF_TOKEN is required to run the text pipeline. Everything else degrades gracefully.

Without HF_TOKEN, the system uses a deterministic Telugu advisory based on ICAR lookup data. It still gives correct, safe answers — just not generative ones.

---

Fine-tuning

The training pipeline fine-tunes google/gemma-2-2b-it on Telugu agricultural Q&A using QLoRA (4-bit NF4, rank 16).

# Build dataset (ICAR-derived + synthetic, ~260 pairs default)
python -m training.dataset

# Add Claude-synthesised pairs (needs ANTHROPIC_API_KEY)
python -m training.dataset --claude-n 500

# Fine-tune (needs GPU — A100 recommended)
python training/finetune.py --hf-repo your-username/krishivaani-gemma2-2b

After training, set HF_MODEL_ID=your-username/krishivaani-gemma2-2b in .env.

---

Evaluation

# Run all evals
python -m eval.run_all

# Individual
python -m eval.safety_eval
python -m eval.latency_eval --runs 10
python -m eval.accuracy_eval

Results land in eval/results/.

---

Project Structure

krishivaani/
├── agent/
│   ├── asr_node.py            # Whisper ASR, confidence-gated
│   ├── router_node.py         # Keyword intent classifier (7 intents)
│   ├── mcp_tools_node.py      # Weather + mandi price live tools
│   ├── llm_node.py            # Gemma-2-2B + deterministic fallback
│   ├── icar_verification_node.py  # Safety layer ← the key contribution
│   ├── memory_node.py         # SQLite session memory
│   ├── tts_node.py            # Edge TTS → gTTS → pyttsx3 fallback chain
│   ├── graph.py               # LangGraph 8-node pipeline
│   └── state.py               # AgentState TypedDict
├── data/
│   ├── icar_lookup.json       # 20 chemicals, ICAR dosages, banned list
│   └── telangana_districts.json  # District → lat/lon for weather
├── training/
│   ├── dataset.py             # Dataset builder (ICAR + synthetic + Claude)
│   ├── finetune.py            # QLoRA training script
│   └── data/                  # train.jsonl, eval.jsonl
├── eval/
│   ├── safety_eval.py         # ICAR verification accuracy
│   ├── latency_eval.py        # End-to-end pipeline latency
│   ├── accuracy_eval.py       # ROUGE-L against held-out set
│   └── results/               # Saved JSON + summary
├── web/index.html             # Browser test UI
├── main.py                    # FastAPI app, 5 endpoints
├── config.py                  # All settings via env vars
└── paper/outline.md           # ACL/EMNLP system paper draft

---

Limitations

• ICAR lookup covers 20 chemicals. CIB&RC registers 800+. Coverage is narrow.
• ASR confidence scoring is heuristic, not derived from Whisper token probabilities.
• No real field deployment yet — evaluation is on curated test sets.
• Fine-tuned model not yet published. Current LLM path is rule-based.
• Telangana districts only for weather geolocation.

---

Roadmap

• Replace heuristic ASR confidence with IndicWhisper + real logprobs
• Publish fine-tuned Gemma-2-2B to HuggingFace Hub
• Expand ICAR lookup to 200+ chemicals
• Telephony integration (Exotel) for actual 2G voice calls
• Cross-session farmer profile (crop, location, disease history)
• Field study with Telangana farmer groups

---

Citation

If you use KrishiVaani in research, please cite:

@misc{krishivaani2026,
  title   = {KrishiVaani: A Safety-Grounded Voice AI Agent for Telugu-Speaking Farmers},
  author  = {Dhamacharla, Asheesh},
  year    = {2026},
  url     = {https://github.com/asheesh07/krishivaani}
}

---

Acknowledgements

• AI4Bharat for IndicWhisper and Bhashini TTS
• ICAR-NIPHM for chemical safety advisories
• LangGraph for the pipeline framework
• Bhashini / MeitY for the Indic language AI infrastructure

---

Built to close the agricultural information gap for low-literacy, low-bandwidth, low-income farmers in India.