Pick a question and press Play. Watch it pass the guards, check the cache, get routed by our own classifier, walk the service graph, get written live by GPT-5.6 Luna when needed, and come back as sourced answer blocks, with the time and rupee cost of every step.
Frontend prototype · keralam.offware.in ↗This page · keralam-engine.offware.in
Cache, then graph templates, then one low-cost model. About 6 in 10 questions never reach the writer; the rest are composed live by GPT-5.6 Luna.
The verifier checks each block cites a retrieved page, and that fees and dates match it word for word.
Emergencies, party politics and advice limits are handled by rules and our own router. Multi-step chains are walked in the graph, not reasoned by the model.
Only anonymous metadata and a redacted question are logged. Files are read in memory and dropped.
Illustrative timings and costs, in rupees at ₹96 per US$ (30-09-2026). GPT-5.6 Luna per million tokens: ₹19 input, ₹1.9 cached input, ₹115 output. Own classifier, embeddings and reranker are counted as small fixed costs. Real numbers come from the pilot.