Book A Readiness Assessment
Vaman Models

Your model. Trained on your portfolio. Benchmarked on your data.

A general-purpose AI model has never read your loan agreements, heard your borrowers, or learned what repayment patterns look like in your geography. Vaman trains a model on what it needs to know — your portfolio — and benchmarks it before it speaks to a single borrower.

Request your assessmentSee how we benchmark
The case for a smaller model

General models are optimised for everything. That means they're not optimised for you.

Domain knowledge

A model trained on your LMS records, call transcripts, and policy documents understands the language of NBFC lending — DPD buckets, EMI schedules, RBI circulars — without ambiguity.

Conduct control

A smaller, fine-tuned model is easier to constrain. You define what it can and cannot say. Guard enforces those rules. A general frontier model is harder to govern with the same precision.

Cost at scale

A focused model running locally costs a fraction of what frontier API pricing charges at the volume an NBFC needs for daily servicing and collections.

Auditability

You can inspect a model trained on your data. You know what went in. For an RBI examination or an internal board review, that transparency matters.

Three tiers

Start where you are. Scale as you grow.

Every institution starts from a different point. Ground, Tune, and Forge match your data readiness and ambition.

Ground

Start with what works.

Vaman base models, adapted to your institution type and regulatory context. Trained on a broad NBFC lending corpus and fine-tuned to follow your calling hours, language rules, and product vocabulary. The fastest path to a live agent.

Good for

First use case
Tight timelines
Limited historical data

What you get

Adapted base model weights — yours to keep
Benchmarked on agreed test cases before go-live
Deployed in your environment
Most commonTune

Your portfolio. Your language. Your model.

A model fine-tuned on your own historical data — your LMS records, call outcomes, payment patterns, and borrower responses. The model learns what works for your specific book, your customers, and your geography.

Good for

Institutions with 12+ months of clean LMS data
Multiple products or geographies
Higher-volume use cases (collections, servicing)

What you get

Fine-tuned model weights — yours to keep
Training conducted inside your environment; data never leaves
Benchmark scorecard: your test cases, your pass threshold
Deployed in your cloud or on-premise
Forge

Built from the ground up for your institution.

A purpose-built model for institutions with complex requirements — multiple loan products, regional languages, custom workflows, or a requirement to own and evolve the model entirely in-house.

Good for

Large or complex NBFC books
Regional language depth across multiple states
Institutions building internal AI capability

What you get

Custom-architecture model weights — yours to keep
Full training pipeline handed over to your team
Joint benchmark design with your risk and compliance teams
On-premise or private cloud deployment
Ownership

When we train a model for you, it becomes yours.

The model weights trained on your data belong to your institution. Your data is never used for another client's model. The Vaman platform — Route, Studio, Ledger, Watch, Guard — is licensed to you separately.

You ownWe license to youWe do
Your dataThe Vaman Stack (Route, Studio, Ledger, Watch, Guard)Build: agents and integrations
Your trained model weightsPlatform updates every quarterRun: managed MLOps, retraining, support
Every output and audit recordReusable agent templates and benchmark toolsAssure: benchmarking, security testing, audit preparation

If you leave, you keep your model and data. The platform is covered by code escrow.

Benchmarks

We prove the difference before you commit.

Every model is benchmarked on a test set agreed with your team before go-live. You set the pass threshold. We do not deploy until it is met.

Task accuracyDoes the agent complete the right action for each scenario in your test set?
Instruction followingDoes the agent stay within your approved content, calling hours, and policy rules?
Escalation judgementDoes the agent route edge cases — disputes, hardship, abuse — to a person?
Language qualityIs the output grammatically correct and appropriate in each supported language?
LatencyIs the response time acceptable for live voice and WhatsApp interactions?

Request a benchmark on your data

We'll plan the benchmark setup with your team before we start.

No data leaves your environment during a benchmark. We agree the setup with your team first.

Questions about models

Do you need access to our production systems to train?

No. Training happens inside your environment. We agree a secure data-sharing arrangement with your IT team for the review period. Nothing leaves your infrastructure.

Who sets the benchmark pass threshold?

You do, in collaboration with your risk and compliance team. We propose a starting scorecard based on our experience; you adjust it to match your internal standards. We do not go live until the agreed threshold is met.

What happens to the model if we leave Vaman?

The model weights are yours. You keep them. The Vaman platform (Route, Studio, Ledger, Watch, Guard) is covered by code escrow — your legal team can review the arrangement before you sign.

Can we upgrade from Ground to Tune later?

Yes. Most institutions start with Ground to get a live agent quickly, then invest in Tune once they have 12 months of production data. The Vaman platform does not change; only the model weights are replaced.

Find out which tier fits your institution.

The readiness assessment identifies which model tier is right for your data, your use case, and your timeline.

Book A Readiness Assessment