Enterprise GenAI Engineer with 4+ years designing and deploying production-grade LLM systems for regulated enterprise workflows. I build AI-assisted finance and compliance systems with human-in-the-loop controls, auditability, and cost-aware architecture — all backed by a strong cloud and backend foundation.
Currently at Jio Platforms (Reliance) shipping LLM-assisted financial decision workflows with deterministic fallbacks, RAG over policy documents with role-based access, and a centralized LLM evaluation/governance framework with prompt versioning, drift detection, hallucination guards, and rollback.
Deep focus on retrieval quality (hybrid search + reranking), structured LLM I/O (JSON mode, Pydantic), trace-level observability (LangSmith / Langfuse-style), and semantic-cache-driven cost control.
Limited.
Designed and deployed enterprise-grade backend systems for finance, procurement, and corporate services workflows, improving response times through architectural and performance optimizations.
Built automated financial decision workflows integrating ML-assisted classification with deterministic validation rules and SAP-based approval systems — auditable, human-in-the-loop processing for compliance-sensitive expense proposals.
Architected a scalable financial management platform that reduced procurement cycle times, with extensibility for future commercialization and strict data integrity guarantees.
Developed event-driven services using RabbitMQ and Node.js to power real-time dashboards and notifications, improving information delivery while maintaining system reliability.
Shipped production GenAI surfaces — a RAG-grounded policy assistant and a centralized LLM Ops platform — to internal users, integrating prompt versioning, hallucination guards, and trace-level observability into the standard enterprise release lifecycle.
Implemented RBAC across internal platforms enforcing strict data boundaries, and collaborated cross-functionally to deliver secure, scalable systems used company-wide.
Enterprise Financial Policy Assistant
- Designed an LLM-assisted financial analysis system grounded in internal policy documents using RAG, constraining responses to approved enterprise knowledge.
- Implemented role-based document retrieval, metadata filtering, and structured prompt templates to prevent unauthorized access and reduce hallucinations.
- Improved retrieval quality with hybrid search (BM25 + dense vectors over pgvector) and a reranking stage; tuned recursive + semantic chunking to keep grounding tight on adversarial finance queries.
- Streamed structured JSON outputs (Pydantic-validated) directly into the approval console via SSE, with token-level streaming for sub-second perceived latency on long answers.
- Introduced confidence-based routing and human-in-the-loop approval flows with full audit trails of prompts, retrieved context, model outputs, and decisions.
- Evaluated reliability through curated test cases and production feedback, monitoring correctness, latency, and token usage to optimize cost-performance.
LLM Evaluation, Monitoring & Governance
- Built a centralized framework for testing, monitoring, and governing LLM behavior across enterprise applications.
- Implemented prompt and configuration versioning, treating GenAI behavior as code with traceable changes and rollback capability.
- Designed automated response-quality scoring pipelines using LLM-as-a-Judge rubrics (correctness, faithfulness, policy compliance) plus heuristic checks to detect hallucinations, retrieval failures, and drift.
- Instrumented end-to-end trace logging across prompt → retrieval → generation → output (LangSmith / Langfuse-style spans) for post-hoc replay and root-cause analysis on failed runs.
- Monitored latency, token usage, and error rates with semantic caching and per-tenant budgets to enforce cost and reliability budgets for production GenAI systems.