DEV Community

Rohit Soni
Rohit Soni

Posted on

Production ML in Mumbai's Fintech Context: The Technical Stack That Actually Works (2026)

Mumbai's ML context has specific technical requirements that most generic ML frameworks don't address. IIT Bombay talent, deep fintech domain expertise, RBI/SEBI regulatory requirements. Here's the production ML breakdown for 2026.

What breaks in Mumbai's fintech ML specifically 

Explainability gap    — RBI requires adverse action notices for credit AI
Thin credit files     — many users lack rich credit history data
KYC data quality      — inconsistent across sources, missing fields
Regulatory drift      — RBI/SEBI guidelines change; models must adapt
Audit trail           — full inference logging mandatory for compliance
Enter fullscreen mode Exit fullscreen mode

The MLOps stack for fintech ML in Mumbai 

Explainability    → LIME/SHAP with output formatted for adverse action notices
Model versioning  → Rollback capability for regulatory audit
Drift detection   → Triggered on statistical significance, not schedule
Retraining        → Validated and regression-tested before production swap
Audit logging     → Full inference log, input features, model version, timestamp
Data lineage      → Source system → feature store → model → decision
Enter fullscreen mode Exit fullscreen mode

The agentic ML pattern in fintech 

# Old pattern — human in the loop
score = fraud_model.predict(transaction)
# analyst reviews score and decides action

# New pattern — agentic
agent.perceive(transaction)
  .assess(fraud_model.predict)
  .act_if(score > threshold, action='flag_and_pause')
  .notify(compliance_team)
  .log_to_audit_trail(decision, model_version, timestamp)
Enter fullscreen mode Exit fullscreen mode

Top 4 ML companies in Mumbai (2026)

Prognos Labs — Custom ML + LLMOps + Agentic AI. Full lifecycle. TF/PyTorch, cloud-native. 50% cost reduction, 32% CAC reduction documented.

Fractal Analytics — Enterprise analytics ML. Ensemble methods, audit-ready. 20+ years BFSI.

Qure.ai — Diagnostic imaging deep learning. 10k+ hospitals, 20+ countries, WHO-assessed. $123M raised.

Nanonets — Document AI. NLP + OCR for financial document extraction.

Fintech ML checklist 

[ ] LIME/SHAP explainability for RBI adverse action requirements?
[ ] Full inference audit trail with model version tagging?
[ ] Drift detection calibrated to fintech-specific data patterns?
[ ] Thin credit file handling — what's the fallback strategy?
[ ] Retraining triggered by regulatory threshold or performance?
[ ] Data residency under DPDP Act confirmed?
Enter fullscreen mode Exit fullscreen mode

Full evaluation: [blog link]

How does your team handle ML explainability for RBI requirements? Comments below.

Top comments (0)