Akhada is a working civic-deliberation engine that runs biographically-grounded AI personas — drawn from real Indian demographic distributions — through a structured debate on any policy topic, synthesises a sourced article + Bradley-Terry quality-weighted conclusive remark, and persists every debate with a cryptographically verifiable hash-chained audit trail.
This is the first cite-able release — pinned in the methodology paper and in the pilot offer.
A SARGVISION Intelligence product.
Documentation set
| Document | Audience |
|---|---|
| README.md | Engineers — 16-section technical reference |
| docs/RELATED-WORK.md | Reviewers / academic — 10-section literature survey, comparison matrix, eight-point gap analysis, 33 references |
| docs/PROJECT-PROPOSAL.md | Programme officers / grant boards — 18-month research-and-implementation proposal, ₹3.0–4.5 cr indicative budget |
| docs/BUSINESS-PROPOSAL.md | Multilateral programme officers / think-tank directors — strategic doc with ₹4 lakh pilot offer + corrected ₹29,008 cr market sizing |
| docs/README.md | Navigation index |
What ships in v0.10.0
| Layer | Capability |
|---|---|
| Persona library | 50 personas spanning 27 states + UTs · 5 hand-curated + 45 Pro-tier-generated · biographically rich (top_5_books with kind discriminator for non-literate personas) · validated against Open Library |
| Sampling | Exact k-DPP (Kulesza-Taskar 2012 spectral algorithm) on TF-IDF embeddings · provable negative-association property |
| Debate runtime | Async Gemini 2.5 Flash R1 fan-out → Pro 7-section synthesis → Pro Bradley-Terry-leaning conclusive remark |
| Persistence + audit | SQLite store · 4-event hash-chained audit log per debate · executable tamper-detection test |
| API | POST /v1/debates · GET /v1/debates · GET /v1/debates/{id} · GET /v1/debates/{id}/audit · GET /v1/runtime |
| Studio | Next.js 14 + Tailwind + Source Serif 4 · light + dark · drop-cap article · audit-chain visualisation page |
| Tests | 31 / 31 passing in 0.80 s |
Defensible position vs. surveyed comparables
Per docs/RELATED-WORK.md §7, Akhada is the only system in the surveyed landscape (Smallville · AI Town · Concordia · Habermas Machine · SYNTHIA · Pol.is · Decidim · Remesh · Anthropic CCAI) that ships all three of: Indian demographic grounding · provable-diversity sampling (k-DPP) · hash-chained audit trail.
Quickstart
git clone https://github.com/sargupta/akhada-engine
cd akhada-engine/services/orchestrator
poetry install
poetry run pytest -q # 31 passed
poetry run uvicorn akhada.api.main:app --port 8080For online mode (real Gemini), see README.md §2.3.
License
AGPL-3.0-or-later (open core) + commercial license issued by SARGVISION Intelligence for gov-tier features. See LICENSE and NOTICE.
Cite this release
Akhada Open Debate Engine, v0.10.0 (May 2026). SARGVISION Intelligence. https://github.com/sargupta/akhada-engine/releases/tag/v0.10.0