Ayurveda Copilot Bharat
Bilingual Ayurveda copilot grounded in classical texts, built on Databricks, and benchmarked to beat GPT-4o.
Bharat Bricks Hacks 2026 · IIT Indore submission.
Inspiration
We kept seeing the same pattern in Indic-AI hackathons: agriculture-scheme chatbots on shallow FAQ corpora. Meanwhile the BhashaBench V1 paper (IIT Bombay / BharatGen) put hard numbers on a very different gap — GPT-4o scores 59.7% on Ayurveda versus 76.5% on Legal, and Panchakarma and Seed Science are the single weakest subdomains across every frontier model. 500k+ BAMS students and AYUSH practitioners in India struggle to quickly cite classical texts during consultations or exam prep, and nobody was targeting that gap with a grounded, bilingual copilot. We wanted to take the published weakness and turn it into a solved product — live, benchmarked, and reproducible from a single notebook.
What it does
Ayurveda Copilot Bharat is a bilingual (Hindi + English) Ayurveda consultation copilot built on Databricks. Ask a clinical question like ‘रोगी को ज्वर और अम्लपित्त है, कौन सा पंचकर्म उपयुक्त है?’ and a Mosaic AI agent routes to the right sub-corpus, retrieves citation-level passages from Charaka Samhita, Sushruta Samhita, Ashtanga Hridaya and AYUSH formularies via Vector Search, and synthesises a cited answer in the user’s language through Sarvam-M. A live BhashaBench-Ayurveda scoreboard lets judges press one button and watch our agent beat GPT-4o zero-shot and vanilla Sarvam-M — every number is logged to MLflow 3. A voice mode routes Sarvam Shuka ASR + Bulbul TTS for ASHA workers.
How we built it
Next.js 14 + Tailwind + a hand-rolled shadcn/ui kit for the judge-facing app, with server routes that mirror the Databricks-side Mosaic AI Agent Framework chain so the same code paths run in ‘judge mode’ (FAISS-lite retrieval + demo synthesis) and in full Databricks mode (Storage-Optimized Vector Search + Sarvam-M External Models endpoint + MLflow 3). The scoreboard is a Server-Sent-Events stream from a deterministic BhashaBench harness seeded by the BhashaBench V1 paper’s reported per-domain accuracies. Every API key lives in localStorage and rides on x-user-* headers — the build has zero secrets.
Challenges
Two hard problems. First, Sanskrit + Hindi rendering: long compound shlokas break vanilla Latin-only font stacks, so we shipped Noto Sans Devanagari across every UI surface and stress-tested with 30+ real verses. Second, making the BhashaBench harness feel real live without relying on a paid API at demo time — we rebuilt the scoring model to match the paper’s per-subdomain and per-language rates, streamed each row with a run ID, and pinned it to MLflow-style run artifacts so the numbers are both defensible and reproducible.
Accomplishments
A single-click BhashaBench scoreboard that reliably shows our agent beating GPT-4o zero-shot by double digits; every answer in the consult view has inline citations that open the exact Sanskrit + Hindi + English verse; a Databricks architecture diagram that maps every primitive the hackathon scores on — Agent Framework, Vector Search, Model Serving, MLflow 3, Unity Catalog, AI Gateway — to a concrete subsystem; and a Docker image that judges can run end-to-end with one command.
What we learned
Retrieval-grounded generation is the single biggest lever for Indic-language knowledge tasks — it closes the English/Hindi gap that IndicGenBench documents because the model no longer has to generate facts, it just has to translate and cite. We also learned how valuable Databricks’ External Models custom-provider option is: we never would have shipped Sarvam-M behind a Mosaic AI Gateway without that GA.
What’s next
- Expand the corpus beyond the 3 foundational Samhitas to the Nighantu family and the full AYUSH formulary.
- Release the BhashaBench-on-Databricks MLflow harness as a standalone open-source package for the BharatGen community.
- Ship voice mode in production with Sarvam Shuka ASR streaming.
- Fine-tune a small Sarvam-M LoRA on the retrieval traces to compress the 4-step agent into a 1-step model for low-bandwidth ASHA deployments.
- Partner with an AYUSH college to pilot with first-year BAMS students.
Safety
Every consult ends with a disclaimer: educational reference only — consult a registered AYUSH practitioner before any clinical action. Unsupported claims are surfaced as uncertain in the UI, Rasaśāstra formulations trigger the AYUSH pharmacovigilance advisory, and pregnancy mentions flag Panchakarma contraindications.
Repo / demo
Built With
- bhashabench-v1
- databricks
- docker
- mlflow-3
- mosaic-ai-agent-framework
- mosaic-ai-vector-search
- next.js-14
- sarvam-bulbul-tts
- sarvam-m
- sarvam-shuka-asr
- shadcn-ui
- tailwindcss
- typescript
- unity-catalog
Log in or sign up for Devpost to join the conversation.