Experience
A timeline of research collaborations and industry roles focused on applied AI, LLM systems, and enterprise-grade automation.
AI Intern
• Finetuning an LLM-driven automation framework (browser-use) that independently performs user-specified queries on websites (Looker Studio, Hootsuite, etc) or Google Sheets by generating and executing custom Playwright-based actions.
• Developed an AI-agentic system powered by dynamic graphs and real-time factory operation databases to support supervisor workflows, detect anomalies, and dynamically incorporate worker-input special conditions.
• Evaluated and productionized computer-vision pipelines for retail supermarkets, fine-tuning YOLOv11 multi-class detectors for Xiaomi (0.87 mAP@50 across over 500 stores in India).
• Built legally-compliant speaker-diarization pipelines for sales–customer calls using open-source (SpeechBrain) and commercial (ElevenLabs) stacks to enable accurate speaker labeling and downstream analytics.
Research Intern
• Developed a heuristic framework based on Cognitive Load and Discourse Analysis to optimize peer review lengths using information density, content relevance, argument strength, and readability metrics. (Published in Scientometrics)
• Designed MetaSearch, a search-augmented, reasoning-based AI Agent for consensus resolution in peer review, combining disagreement detection and fact-grounded synthesis to automate meta-review decision-making.
• Developed a novel LLM-argumentation system for peer review, evaluating 900+ ICLR reviews (PeerRead Dataset) using leading LLMs (Llama 3.1, Mixtral, Gemini2, GPT-4o), achieving high inter-annotator agreement (Cohen's Kappa = 0.834).
• Engineered BiasChain, a modular multi-agent LLM framework for justified peer review bias detection, by orchestrating specialized agents for sentiment coherence, internal consistency, and inter-review alignment.
AI Intern
• Authored a whitepaper implementing an evaluation pipeline for LLM-based requirement generation, benchmarking DeepSeek R1, GPT-4o, Gemini 2.0, LLaMA 3.2 across diverse NLP metrics including BLEU, Levenshtein and Jaccard Similarity.
• Devised Retrieval-Augmented Generation (RAG) workflows in Azure AI PromptFlow across 50+ RFPs.
• Optimized chunking strategies and top-p sampling parameters, reducing document retrieval perplexity by 15%.
• Authored a comprehensive 70-page documentation detailing the mapping of AI APIs (Insomnia) to the frontend (NextJS).
Research Assistant
Investigated C-to-Rust transpilation with existing tools and LLMs, categorising translation bugs and strengthening quality evaluation frameworks for future research.
Python (LLM) Development Intern
Built AI-powered marketing content pipelines, evaluated multiple image-generation models, and tailored GenAI prompts to create platform-specific, brand-aligned campaigns.