Thiruvaasal
Probabilistic Reconstruction of Lost Pallava Temples
What It Is
A research system for reconstructing the six lost Pallava temples of Mahabalipuram using evidence-tiered inference, Agamic architectural constraints, posterior reweighting, and per-element uncertainty maps. The system does not claim a single historical truth — it characterizes the posterior over plausible reconstructions given the evidence.
Most AI-driven 3D reconstruction is unconstrained generation — it hallucinates spires, invents floor plans, and produces nothing a heritage reviewer can verify. Thiruvaasal asks whether canonical architectural priors can make archaeological reconstruction *falsifiable, calibrated, and reviewable*. Constraint necessity is tested by ablation, not asserted.
What I Built
- Evidence ledger with explicit tier / source / rights / geometry scope / provenance per element
- Agamic hard-constraint engine — rejects impossible structures before scoring; soft constraints scored as named energy terms with visible weights
- Posterior reweighting pipeline producing per-element confidence from posterior marginals
- Holdout temple experiment design — tests the method on a temple with known ground truth before generalizing
- Constraint necessity ablation — quantifies whether priors are doing real work or are decorative
- Reviewer-ready report generation with full audit trail (config + input artifacts + output artifacts + verification command per result)
Tech Stack
Phase 1 method in active development. Paper bet is narrow: one site + one holdout protocol proving the priors before any public demos. Foundation for the KUMARI multi-site knowledge OS.