BharatAI — Unified Legal & Health AI Assistant

India has 1.4 billion people, most of whom cannot afford a lawyer or a doctor. BharatAI bridges that gap — a unified AI assistant that answers Indian criminal law questions under BNS 2023 and rural health queries, in Hindi, English, and Hinglish, with full voice support.

How It Works

Under the hood, a Qwen3-8B classifier routes every query to the right pipeline in real time.

Legal Pipeline

Legal queries hit a RAG system built on GTE-Large embeddings over $N = 448$ chunked BNS sections stored in PostgreSQL. Retrieval uses dot-product cosine similarity:

$$ \text{score}(q, c_i) = \frac{\mathbf{e}q \cdot \mathbf{e}{c_i}}{|\mathbf{e}q| |\mathbf{e}{c_i}|} $$

where $\mathbf{e}q \in \mathbb{R}^{1024}$ is the query embedding and $\mathbf{e}{c_i}$ is the chunk embedding. Top-$k$ chunks ($k = 10$) are retrieved, with co-retrieval of offense–punishment pairs. Sarvam Translate handles Hindi $\leftrightarrow$ English conversion; Qwen3-8B generates cited answers with BNS section numbers.

Health Pipeline

Health queries go through a 5-step Sarvam 30B orchestration pipeline:

$$ \text{Query} \xrightarrow{\text{Step 1}} \text{Understanding} \xrightarrow{\text{Step 2}} \text{Search Optimisation} \xrightarrow{\text{Step 3}} \text{Parallel Fetch} \xrightarrow{\text{Step 4}} \text{Synthesis} \xrightarrow{\text{Step 5}} \text{Safety Check} $$

Step 3 fans out to $M \geq 10$ sources concurrently:

Emergency keywords (chest pain, seizure, snake bite, etc.) set $\texttt{is_emergency} = \texttt{true}$, triggering an immediate red-alert banner with helpline numbers 112 / 108.

Voice Pipeline

$$ \text{Audio (16kHz WAV)} \xrightarrow{\text{Sarvam Saarika v2.5 ASR}} \text{Transcript} \xrightarrow{\text{Router}} \text{Pipeline} \xrightarrow{\text{IndicF5 / Bulbul v2 TTS}} \text{Audio response} $$

• STT: Sarvam Saarika v2.5 ASR (hi-IN, en-IN)
• TTS (Hindi, primary): IndicF5 — open-source, voice cloning via HuggingFace Space
• TTS (fallback): Sarvam Bulbul v2

Deployment & Data Architecture

The full stack — FastAPI backend + React SPA — is deployed as a single-process app on Databricks Apps. The BNS data pipeline follows the medallion architecture built in Databricks Notebooks:

$$ \underbrace{\texttt{bns_raw_sections}}{\text{Bronze: raw CSV}} \;\longrightarrow\; \underbrace{\texttt{bns_sections},\; \texttt{bns_section2_definitions}}{\text{Silver: normalised}} \;\longrightarrow\; \underbrace{\texttt{bns_chunks},\; \texttt{bns_chapter_summaries}}_{\text{Gold: embeddings } \in \mathbb{R}^{1024}} $$

Models & Open-Source Stack

Built With

• aiohttp
• css3
• databricks-notebooks-databases-postgresql-15-with-pgvector-(1024-dim-hnsw-index)-ai-/-ml-models-qwen3-8b
• google-custom-search-api
• gradio-client-platforms-databricks-apps
• html5
• httpx
• huggingface-inference-api
• huggingface-spaces-cloud-services-azure-postgresql-(or-aws-rds)
• indicf5
• javascript-(es2023)
• languages-python-3.11
• medlineplus-xml-api
• numpy
• openfda-api-other-web-audio-api-(browser-side-wav-recording)
• pubmed-e-utilities
• pydantic-v2
• react-18
• react-markdown
• sarvam-30b
• sarvam-bulbul-v2-tts
• sarvam-saarika-v2.5-asr
• sarvam-translate-apis-sarvam-ai-api
• sentence-transformers
• soundfile
• sql-frameworks-&-libraries-fastapi
• tenacity
• thenlper/gte-large
• vite-5
• youtube-data-api-v3