Our proposalWhat we can do
Our proposal · what we can do

Custom models and inference, on one context.

Four things we build with Policybazaar. The first two are the core.

  • 1Custom trained modelsCore
    Beat frontier models at 1/60th the cost
    Our Banking77 study: a post-trained 0.5B model, 88.1% vs 81.2% for Claude Sonnet 4.5, ₹1.30 vs ₹83 per 1,000

    PB Flash and PB Pro, trained on your data.

    • Trained on your chats, calls and outcomes
    • Released only when they beat the frontier on your evals
    • Owned by Policybazaar, at a fraction of the cost
  • 2InferenceCore
    Save up to 85% on inference spend
    pb.ai's 31 test chats: ₹0.13 a chat on open models vs ₹0.99 on today's Gemini stack, at higher quality

    The right model for each job, served in India.

    • One API for every agent
    • The cheapest open model that passes your tests
    • A custom model where it pays
  • 3Shared context layer
    Zero questions asked twice
    pb.ai's test chats: 5 of 5 ask-once checks passed, against 2 of 5 on today's stack

    One memory of each customer, under every channel.

    • Loaded on every turn, on every channel
    • Each fact kept with its source
    • Said once, never asked again
  • 4Evals
    Every answer checked before it ships
    Routing, prices, safety and style, scored on every release

    Prove it works, on Policybazaar's own chats.

    • Each agent tested on Policybazaar's past chats
    • You set the pass mark; nothing goes live below it
    • Every new version re-tested before go-live