I build practical AI/ML systems with a focus on Agentic AI, large language models, retrieval-augmented generation, multilingual speech AI, and AI applications for digital payments.
My work combines hands-on experience in the digital payments ecosystem with building and evaluating AI systems for real-world applications, particularly across financial use cases, Indian languages, and intelligent automation.
- π€ Agentic AI & LLM Orchestration
- π§ Large Language Models (LLMs)
- π Retrieval-Augmented Generation (RAG)
- π οΈ LLM Fine-Tuning & QLoRA
- π£οΈ Speech AI & Automatic Speech Recognition (ASR)
- π Multilingual AI for Indian Languages
- π Tool Calling & AI Workflow Automation
- π‘οΈ AI Guardrails & Groundedness
- π Model Evaluation & Benchmarking
- π AI Inference & Model Serving
- π³ AI for Digital Payment Systems
An agentic AI system for UPI payment operations and failure investigation, combining structured transaction analytics, operational RAG, grounded reasoning, validation/retry workflows, human-in-the-loop approval, and simulated incident management.
Highlights:
- LangGraph-based agent orchestration
- Planner and intent-based workflow routing
- SQLite-based payment transaction analytics
- Operational RAG using ChromaDB and Sentence Transformers
- Local LLM inference with Ollama
- Evidence-grounded investigation generation
- Guardrails for unsupported numeric and causal claims
- Validation and retry workflow
- Human-in-the-loop approval before incident creation
- Simulated incident management
- MCP server exposing operational tools
- Automated evaluation framework
- 37 automated tests
- Synthetic evaluation datasets for intent, retrieval, groundedness, task completion, and recovery
π£οΈ IndicConformer ASR Benchmark
Benchmarking AI4Bharat IndicConformer-600M for multilingual Automatic Speech Recognition across 9 Indian languages and 900 audio samples.
Highlights:
- 9 Indian languages
- 900 audio samples
- Word Error Rate (WER) evaluation
- IndicConformer-600M
- RNNT decoding
- CPU vs GPU inference benchmarking
A Retrieval-Augmented Generation system for answering questions from RBI regulatory documents using semantic retrieval and local LLM inference.
Highlights:
- PDF document processing
- Semantic chunking experiments
- Sentence Transformers embeddings
- ChromaDB vector search
- Local LLM inference with Ollama
- Retrieval evaluation using Recall@1, Recall@5 and MRR
Fine-tuning Microsoft Phi-3 Mini 3.8B for payment-related intent classification.
Highlights:
- Phi-3 Mini 3.8B
- 7 payment intents
- QLoRA
- 4-bit quantization
- Parameter-efficient fine-tuning
- Baseline vs fine-tuned evaluation
- Python
- SQL
- PyTorch
- Hugging Face Transformers
- Sentence Transformers
- Large Language Models
- RAG
- Fine-Tuning
- QLoRA
- LangGraph
- LangChain
- Tool Calling
- Agent Workflows
- AI Guardrails
- Human-in-the-Loop
- MCP
- ChromaDB
- SQLite
- Ollama
- Docker
- Git
- Linux
I am particularly interested in building AI systems that move beyond experimentation into reproducible, measurable, and production-oriented solutions.
My current areas of focus include:
- Agentic AI systems
- LLM applications
- Retrieval systems
- Efficient LLM fine-tuning
- Multilingual Speech AI
- AI evaluation and benchmarking
- Grounded generation and guardrails
- AI inference and model serving
- AI applications for digital payments
β Feel free to explore my repositories to see the implementations, experiments, evaluations, and engineering decisions behind these projects.