AI Engineer | ML Engineer | Applied AI
LLMs · AI Agents · RAG · Machine Learning · Computer Vision · Multimodal AI
I build and evaluate AI systems that can reason, retrieve, perceive, and act. My work spans LLM engineering, AI agents, retrieval-augmented generation, machine learning, and multimodal systems, with experience across model architecture, training, evaluation, and end-to-end system design.
I hold an MSc in Artificial Intelligence from Queen Mary University of London and a BTech in Computer Science and Engineering from Vellore Institute of Technology.
I’m particularly interested in turning advances in AI research into reliable, scalable systems that can solve complex real-world problems.
I’m always keen to connect, exchange ideas, and explore new conversations and opportunities across AI and technology.
London, United Kingdom
Python, Java, C/C++, SQL
PyTorch, Hugging Face, LangChain, LangGraph, CrewAI, BMAD Method, OpenCV, spaCy, Spring Boot
Cursor, Claude Code, Git, GitHub, Jupyter, Docker
PostgreSQL, Qdrant, Amazon Web Services
A hybrid RAG pipeline over the SciFact corpus: BM25 first-stage retrieval, ColBERT-style late-interaction re-ranking, and citation-backed generation from a 4-bit quantised Qwen2.5-3B model. The ColBERT projection layer was fine-tuned with InfoNCE on BM25-mined hard negatives.
| nDCG@10 | AP@10 | RR@10 | Re-ranking latency |
|---|---|---|---|
| 0.6858 | 0.6479 | 0.6581 | 43.1 ms / query |
A five-agent pipeline that produces investment reports from a natural-language ticker or cryptocurrency name. Components cover identifier resolution, financial data retrieval, news sentiment, and a ten-year discounted cash-flow model with WACC-based discounting. Retrieval over SEC 10-K filings uses Qdrant, and reports are generated with ReportLab.
A lightweight multimodal retrieval system that maps images to semantically related text labels without a large pretrained foundation model. A Skip-Gram backbone trained on Visual Genome was expanded with a (1+λ) evolutionary algorithm to incorporate CIFAR-100 labels, then aligned to MobileNetV3 features with InfoNCE. Image-to-text Recall@10: 91%.
An agentic system that generates recipes from a photograph of ingredients or a prepared dish, combining CLIP, Qwen2.5-3B, FAISS retrieval, Tavily, and Llama 3.
Designed ReVA, a region-aware VQA model that combines global scene context with localized visual evidence to improve fine-grained visual grounding and reduce hallucination.
ReVA connects a frozen CLIP ViT-L/14 vision encoder to Qwen2.5-7B-Instruct through separate image and region bridges. Object proposals from RAM++, spaCy, and Grounding DINO supply bounding boxes. RoI-aligned intermediate ViT features are mapped to region tokens so that answers are conditioned on local visual evidence as well as global scene context.
| VQAv2 | MMBench | SEED-Bench | POPE |
|---|---|---|---|
| 73.79% | 67.1% | 66.75% | 82.85 F1 |
To read the paper or download the implementation and weights, use the links below.
| Sep 2025 – Sep 2026 |
Master of Science, Artificial Intelligence
Queen Mary University of London, London, United Kingdom
|
|---|---|
| Jul 2021 – Jul 2025 |
Bachelor of Technology, Computer Science and Engineering
Vellore Institute of Technology, Vellore, India
|
| Jan 2025 – May 2025 |
AI Research and Development (R&D) Intern
DRDO, Ministry of Defence, Government of India, Bangalore Developed an Object Diagram Recognition system using Computer Vision and Deep Learning to interpret control-law diagram images and automatically generate executable C code. Fine-tuned YOLO and integrated efficient line-detection algorithms using OpenCV to extract precise component bounding boxes and control-flow coordinates from noisy diagram images. Constructed logical representations of diagrams as directed acyclic graphs (DAGs) and applied graph-parsing techniques for automated code generation, reducing manual coding effort and syntax errors in complex control-law diagrams. |
|---|---|
| Aug 2023 – Nov 2023 |
Software Developer Intern
Pivotrics, Bangalore Developed scalable RESTful APIs for a SaaS-based payment platform to enhance customer experience, utilizing Spring Boot, WebFlux, and PostgreSQL. Optimized system performance during peak load by adopting reactive programming principles, designing non-blocking, asynchronous APIs with effective back-pressure handling for seamless management of product information. Implemented JWT-based authentication and authorisation with robust access control policies using Spring Security. |