• Contributed to a production-grade agentic B2B outreach system integrating email sending, receiving, replying, and campaign-level tracking across multiple mailboxes.
• Designed and implemented the LLM tool layer (send/track/reply) with persistent conversation state in PostgreSQL, enabling autonomous, multi-threaded communication workflows.
• Engineered a real-time feedback pipeline capturing 16+ delivery and engagement signals, enabling adaptive, data-driven campaign optimization.
• Developed a context-aware RAG pipeline leveraging prospect data for personalized email generation, improving relevance over template-based outreach.
• Refactored the FastAPI backend to resolve critical issues and deployed services on Azure Function Apps, enabling a scalable, event-driven system.
• Built multi-stage RAG pipelines and agentic workflows to simulate real-world LLM usage and enable systematic evaluation of faithfulness, relevance, and failure modes in complex pipelines.
• Designed advanced retrieval pipelines (query routing, HyDE, reranking) to generate challenging, multistep queries, improving robustness and realism of evaluation benchmarks.
• Reverse engineered 3 LLM observability frameworks to understand tracing architectures and extract design insights for building in-house evaluation and observability tooling.
• Built a PoC for an in-house LLM tracing system to validate design decisions and inform the product team in developing a production-grade tracing SDK.
• Designed and validated a data-free, zero-shot machine unlearning framework to enable class-level data removal without requiring access to original training data, supporting privacy compliance.
• Leveraged contrastive model inversion to generate 5K synthetic samples, enabling effective targeted unlearning while preserving model performance on unaffected classes.
• Achieved 90%+ unlearning efficacy across 8 classes on the SVHN dataset, while maintaining 95%+ accuracy on retained classes, minimizing unintended knowledge loss.
• Improved unlearning efficiency with a 77% reduction in runtime compared to the GKT baseline, enhancing feasibility for real-world deployment.
• Conducted comparative evaluation against 4 baseline methods (EMMN, GKT, zMuGAN, retraining) and explored subset/coreset strategies to optimize scalability and effectiveness of unlearning.
AI / ML / DL
LLM Systems
Backend & Data
Cloud & Infra
Tracking
MSc Data Science
DIAT, PuneBSc Mathematics
IGNOUEnterprise-grade Multi-Tiered Retrieval-Augmented Generation System. Integrates dynamic fallback strategies and diverse retrieval architectures to ensure high-fidelity context extraction across complex datasets.
Custom-built, local-only LLM tracing and observability tooling. Engineered for async agentic workflows using contextvars to provide granular telemetry, span tracking, and metric evaluation with minimal overhead.
Multi-path autonomous RAG pipeline featuring intelligent query routing, self-corrective retrieval loops, and iterative web search fallbacks to actively detect and eliminate hallucinations in real time.
On-device RAG engineered around LEANN's graph-based search. Indexes millions of documents locally in gigabytes without significant accuracy loss, achieving zero cloud dependency and strict privacy.
Structure-aware documentation RAG pipeline built from scratch. Crawls 104 FastAPI pages, preserves section hierarchy, generates 971 metadata-rich chunks, and delivers citation-grounded answers via Pinecone + Groq.
RAG system with SHA-256 document deduplication, namespace isolation, semantic reranking, and inline-cited answers. Upload a 200-page manual, ask a question, get a precise cited response.