Solving for India at Scale: Use of AI in Fintech

Abstract

The session opened with a high‑level view of PhonePe’s massive reach—hundreds of millions of users and merchants across every Indian pin‑code—and explained why building AI at this scale demands a “inside‑out” approach. The speakers described PhonePe’s end‑to‑end AI stack, from the on‑premise LLM gateway (named Godric) and a lightweight Java‑based agent framework (Sentinel), to a purpose‑built vector store (Severus) and the open‑source Agent Hub that powers thousands of internal assistants. They then illustrated how AI is being used to streamline developer productivity, fraud detection, merchant onboarding, and a host of consumer‑facing features such as natural‑language search, personalized transaction insights, credit‑score analysis, and AI‑driven support. Throughout, the team emphasised security, compliance, and trust as non‑negotiable pillars of every rollout. The discussion closed with a short audience Q&A covering future fintech‑innovation directions, crypto, and competitive positioning.

Detailed Summary

User & Merchant Base – PhonePe serves ≈650 million registered users (now higher) and ≈47 million merchants. Roughly 300 million users are active each month, sending money to 11‑12 million merchants and generating ≈52‑53 billion transactions (≈$870 billion value) over six months.
Geographic Penetration – Activity spans every pin‑code in India; the app is used for hospitals, auto services, food, groceries, etc. This ubiquity means any new feature touches the entire nation.
Strategic Imperative – Because any release impacts everyone, PhonePe chose to understand the AI problem internally first, then expose it outwardly. This “intelligent systems inside‑out” philosophy ensures the company masters reliability, security, and compliance before setting organisation‑wide AI adoption goals.

2. Building the AI Foundations

2.1. Homogeneous, Highly Reliable Tech Stack

PhonePe operates a single, uniform tech stack across all services. Consistency is essential for reliability at scale.
All internal libraries and infrastructure are built to be well‑tested, production‑grade, and easily reusable.

2.2. AI/ML Platform – The Core Execution Engine

Long‑standing ML capability – PhonePe has deployed batch and real‑time models for years.
The platform handles model onboarding, autoscaling, and lifecycle management for both edge (on‑device) and server‑side models.
Deployment Mechanics – Uses a custom container orchestrator (a tiny two‑package Debian system) that launches stateless Docker containers. No stateful containers; all state lives in external, highly‑available databases.

2.3. LLM Gateway – Godric

Purpose – Provides a unified OpenAI‑compatible chat‑completion API for every downstream application.
Key Features
- Internal OAuth for authentication/authorization.
- Quota Management to protect GPU capacity from runaway usage.
- Safeguard Layer – Dedicated LLMs vet every request for policy compliance and data‑leak prevention.
- Routing Engine – Dynamically directs a request to:
  - On‑prem GPU servers (PhonePe‑hosted),
  - Azure GPU instances, or
  - Third‑party “foundry” models (global or local) depending on latency, cost, and compliance needs.
- Auditing – Every API call (caller, model, payload, timestamp) is logged to a secure, immutable store for regulator and internal audits.

2.4. Agent Framework – Sentinel

Java‑centric to match PhonePe’s engineering culture.
Toolboxes & Extensibility – Allow developers to compose multiple Model‑Control‑Process (MCP) servers, inject custom tools, and truncate responses to stay within LLM context windows.
Agent Registry & Skills – Early precursor to modern “skill” marketplaces; agents advertise metadata that the framework injects into prompts, enabling dynamic skill selection.
Open‑Source Release – QR code provided during the talk; code is publicly available for community contributions.

2.5. Knowledge Store – Severus

Lightweight RAG (Retrieval‑Augmented Generation) engine built on GlusterFS‑based, S3‑compatible storage.
Handles embedding generation, schema‑based document storage, and serves vectors to any LLM via a simple API.
Designed for low complexity to ensure scalability and easy integration with agents.

3. Internal Use‑Cases – Driving Engineer Productivity & Operational Efficiency

Use‑Case	How AI Helps	Outcome
Developer Efficiency	LLMs automate code‑review suggestions, generate boiler‑plate, and assist in refactoring.	Faster feature delivery; reduced cognitive load.
Merchant Document Validation	AI extracts and verifies fields from KYC documents, cutting manual verification time.	Higher throughput for onboarding; fewer errors.
Fraud Investigation Assistant	Preliminary analysis of transaction patterns and red‑flag detection, surfacing only high‑risk cases to investigators.	Reduced investigator workload; quicker response.
Agent Hub (internal)	Centralised GUI to configure agents, connect to MCP servers, and expose them via Slack, email, etc.	≈200‑300 agents in production for a variety of internal workflows.
Clio – Automatic Documentation	Reverse‑engineers source code to produce up‑to‑date design docs.	Saves engineers hours of manual documentation.
Code‑Migration Recipes (Logan‑based)	Suggests language‑level refactors to keep the codebase modern.	Streamlined migration paths, lower technical debt.
QA Test‑Case Generation	Natural‑language prompts produce UI test scripts automatically.	Faster regression testing, higher coverage.
HR Bot	Consolidates leave policies, reimbursement rules, etc., and answers employee queries.	Cuts HR support tickets; improves employee experience.

Key Insight: By embedding AI early in the engineering workflow, PhonePe reduces repetitive toil, allowing engineers to focus on high‑value problem solving while keeping quality and security front‑and‑center.

4. Consumer‑Facing AI Features

Natural‑Language Search (text & voice) across multiple Indian languages.
The system interprets user intent and routes directly to the desired action (e.g., “pay my electricity bill”, “how much did I send my wife last month”).
Hybrid Deployment – Sensitive logic runs on‑device (local inference) while heavy LLMs run in the secure PhonePe data‑center, ensuring zero data exfiltration.

4.2. Transaction Insight Assistant

Users can ask for aggregate spend analytics (e.g., “total bill payments in the last six months”).
The assistant pulls from the user’s transaction ledger, summarises, and presents insights in plain language.

4.3. Credit‑Score Coach

Video Summaries generated by LLMs explain a user’s credit score, actionable tips (e.g., keep utilization < 30 %).
Simulation Engine – Users can model a hypothetical home‑loan scenario and receive a personalized roadmap.

4.4. Customer‑Support AI

Proactive Chatbots that recognise user tone and language, guiding them through new features or troubleshooting.
Escalation Path – When a bot cannot resolve, the case is handed to a human with full context logs.

4.5. Future Consumer Experiments (mentioned briefly)

Voice‑first payments, tokenisation, and “agentic” commerce experiences are under exploration. The team stressed that real‑world usage data will shape the final modality.

5. Security, Privacy & Compliance

Data Residency – All customer PII remains within Indian data‑centres; only non‑sensitive, anonymised data may be sent to global LLM providers for specific internal use‑cases (e.g., HR docs).
Audit Trails – Godric logs who, what, when, where for every LLM call; logs are immutable and searchable for regulator audits.
Model Safeguards – Dedicated LLMs pre‑screen prompts/outputs to prevent leakage of proprietary data or generation of disallowed content.
Authorization – Fine‑grained OAuth validates each service‑to‑service call; quota limits prevent denial‑of‑service attacks.
Compliance – The system is designed to satisfy PCI‑DSS, KYC, AML, and other Indian financial regulations.

6. Open‑Source Contributions & Community Engagement

Godric, Sentinel, Agent Hub, and Severus are all released under permissive licenses.
PhonePe contributes to GlusterFS (maintainers are PhonePe engineers).
The team invited developers to file issues, submit PRs, and use the Mantis security scanner for domain‑level vulnerability checks.

7. Q&A Highlights

Question	Summary of Answer
Scale vs. Innovation – “Will PhonePe adopt emerging infra like Kubernetes or cloud‑native stacks?”	Santanu emphasized that PhonePe builds what works for its scale, preferring a home‑grown orchestrator that guarantees zero‑state loss. The philosophy is “solve the problem first, then adopt the tool”.
Future AI‑driven modalities (voice, tokenisation, etc.)	The team sees AI as a force multiplier but admits the optimal consumer interface (voice vs. chat vs. other) is still unknown; they will iterate based on user behaviour.
Crypto – “How is PhonePe preparing for government crypto initiatives?”	Answer: Early stage – no immediate impact; PhonePe will monitor regulatory developments before integrating any crypto‑related features.
Competitive differentiation – “All payment apps look the same; how does PhonePe stay ahead?”	The response highlighted trust, reliability, and privacy as core differentiators, coupled with continuous AI‑driven user experiences (personalised insights, AI‑assisted support).
Revenue model – “Payments don’t directly generate revenue?”	Clarified that transaction fees, value‑added services, and AI‑enabled cross‑sell do contribute; the business model is robust when solving real user problems.

Key Takeaways

Massive scale requires internal AI mastery. PhonePe built a bespoke, end‑to‑end AI stack (Godric, Sentinel, Severus) before exposing AI capabilities to customers.
Reliability & security are non‑negotiable. All models run in a tightly‑controlled environment with OAuth, quota management, safeguard LLMs, and immutable audit logs.
Homogeneous engineering culture (Java‑first, custom container orchestrator) enables rapid, consistent rollout of AI services across the organisation.
AI is used everywhere: from developer productivity tools (code review, documentation generation) to frontline customer experiences (natural‑language search, transaction insights, credit‑score coaching).
Domain‑specific layers outperform generic tools. After initial experiments, PhonePe invested in custom agents and knowledge stores tailored to fintech use‑cases.
Open‑source mindset – core components are publicly released, encouraging community feedback and positioning PhonePe as a thought‑leader in AI‑at‑scale for fintech.
Compliance built‑in, not an afterthought. Data never leaves Indian borders unless explicitly allowed; every LLM interaction is audited for regulator review.
Future vision: Continue to evolve AI‑driven intent/action flows, experiment with voice/agentic commerce, and maintain a focus on trust, privacy, and inclusive financial access for every Indian.

See Also:

India AI Impact Summit 2026

Explorer

solving-for-india-at-scale-use-of-ai-in-fintech