Overview
Demo · Vercel
Serverless demo: traces and Settings-pasted keys reset on cold start. Put provider keys in Vercel env if you want them to stick. This is not the bank deploy — that is a Node process on their disk.
Use cases

Banks buy journeys, not a model brand. Cost-to-serve and India residency gate vendors. Point production traffic at this box via the gateway — OpenRouter or BYOK in Settings. No keys means bake-offs fail instead of faking scores.

No provider keys. Add an OpenRouter key in Settings (covers Claude / GPT / Gemini) or BYOK. Experiments and /v1 will error until you do.

Spend · eval + gateway
₹0
Gateway calls
0
Gold-set quality
Frontier vs Dvar
Frontier ₹0Dvar ₹0
Digital onboarding
active₹0

Priority cost centre. Scan-and-save and KYC already cut effort; CRO/CISO gate production. Measure cost-to-onboard, not model brand.

Every new-to-bank file · Axis / HDFC

0 eval cells0 prod callsquality p50 0ms
Run this gold set →
Contact centre / servicing
active₹0

60–65% of contact-centre load is the economic prize. +1s latency kills voice. Cost/quality must hold on Hindi + English.

High daily volume · cannot recoup via fees (RBI)

0 eval cells0 prod callsquality p50 0ms
Run this gold set →
RM / wealth assist
active₹0

Policy replies and escalation, not straight-through money movement. Redemptions and waivers stay with a human.

RM desktop · HNI book at risk if the RM leaves

0 eval cells0 prod callsquality p50 0ms
Run this gold set →
Offers / next-best-action
gated₹0

Axis-scale (~50M customers, ~150k/day flows) stays off-LLM until a bake-off shows ~50% cost at ≤5% quality drop. RBI bars charging the customer to recoup tokens.

Highest volume · gated until economics work

0 eval cells0 prod callsquality p50 0ms

Stay off-LLM until a bake-off shows ~50% cost at ≤5% quality drop.

No bake-offs yet. Each playground is one bank use case. Run after you add keys.