The Siddha Lab
Grow a model.
Watch it think.
This is not a video of our product — it is the method, running in your tab. Because Reverse Synthetic Networks are closed-form, a faithful miniature of the embedding-only mode synthesizes below in milliseconds, entirely client-side. Nothing you type is sent anywhere.
Labeled examples
Synthesize the model
One closed-form pass: token–class statistics, IDF, discriminative weighting, class prototypes. No gradients — which is why it can run here, in your tab. Your data never leaves this page.
Classify new text
A faithful miniature of RSN's embedding-only mode — the label-enriched token statistics our research identified as the primary small-scale signal — not the full four-pillar system. The full architecture adds spectral interconnection, synthesized attention, and ensemble convergence on real workloads.
Benchmark explorer
The numbers — wins, losses, nulls.
Three tabs. One is flattering, two are not. All three are the same size, because that is the deal we made with the evidence.
Accuracy, higher is better. The digits row shows the trained MLP ahead — synthesis loses there and we chart it at full size.
Energy estimator
What your cadence costs the grid.
Set how often you'd rebuild a model. The assumptions are printed beside the result, and the scope of the claim beneath it.
High-cadence teams rebuild daily or hourly as data drifts. Drag the slider to your cadence; the arithmetic updates from the stated assumptions below.
Stated assumptions, honestly scoped: synthesis at 60 s / ~35 W (measured order of magnitude for the 20 Newsgroups ensemble); fine-tune at 2 h on one ~400 W A100; grid intensity 0.71 kg CO₂e/kWh (Indian grid average, CEA). The comparison applies to the small discriminative workloads where synthesis is competitive — it is not a claim about LLM training, which synthesis cannot replace.