MSc Computing (AI & Machine Learning) · Imperial College London
I am an MSc student in Computing (Artificial Intelligence and Machine Learning) at Imperial College London. My research is on large language models and natural language processing: how models represent words and meaning, how one prompt can produce several genuinely different answers, and how to make models smaller without losing quality.
Before Imperial I studied Computer Science and Engineering at KL University (CGPA 9.72/10), worked as a research intern at IIIT Hyderabad on layout-preserving document translation, and built integration software for enterprise clients at InfoAjax.
IEEE ICCCMLA 2025 · Published in IEEE Xplore
Every word is built from its characters by a character-level Transformer, so related forms share most of their computation and unseen words still get a vector. In a seeded re-implementation it beats word embeddings by 14.9 F1 on German NER (63.0 vs 48.2) and matches them on English topic classification.
IEEE NEPCON 2026 · Accepted, presentation December 2026
A ViT-Small detector with an edge stream for blending boundaries and a frequency stream for generator fingerprints, fused by iterative cross-attention. After an audit found that most WildDeepfake test frames came from training videos, it reports 96.3% AUC on a video-disjoint test subset (87.1% on Celeb-DF).
IEEE NEPCON 2026 · Accepted, presentation December 2026
Only K seed vectors are trained, added to the input embeddings of a frozen transformer and pushed apart by a repulsion loss, so one prompt yields K different answers in a single batched pass with no filtering. The paper gives the method and formulation. Follow-up (Oct 2026): a redesigned version, tested in pre-registered GPU experiments on Qwen2.5, did not significantly beat tuned min-p sampling; the main suspected cause is the training data.
Zenodo preprint · Version 2, October 2026
A storage format for pruned weights: the gaps between non-zeros are Golomb–Rice coded, close to their entropy limit, and values are 8-bit codes stored as bitplanes. At 95% sparsity it is 44.5× smaller than float32 at 0.6% error, against 25.7× for CSR with the same values.
InfoAjax Consulting · remote
Built a proof-of-concept browser agent that acts on screenshots using models on Azure AI Foundry, and a React and FastAPI console for monitoring agents live.
IIIT Hyderabad
Designed a document translation system that keeps the original layout, for English, Hindi, Telugu and German: Vision Transformer layout detection, font detection, colour extraction and Tesseract OCR. Demonstrated it at the institute’s research expo (demo).
InfoAjax Consulting · remote
Built REST APIs in WSO2 Micro Integrator linking the service-desk, billing and inventory systems of a Philippine telecoms company, and supported them in production.
Imperial College London
KL University, Hyderabad · CGPA 9.72 / 10
iHub-Data, IIIT Hyderabad
Certificate of Achievement — Global Challenge Lab 2026 (1st Place)
TensorFlow Developer Certificate
AWS Certified Cloud Practitioner
AWS Certified Solutions Architect – Associate
Oracle Cloud Infrastructure Architect Associate
Oracle Cloud Infrastructure Generative AI Professional
Oracle Database Certificate
Certified Advanced Automation Professional
Red Hat Certified Enterprise Application Developer
WSO2 Certified Micro Integrator Developer – V4
Build Large Language Models From Scratch
Hugging Face MCP – Unit 1
Hugging Face MCP – Unit 3
IIIT Hyderabad 6-Month AI/ML Training
Certificate of Appreciation — Student Activity Center
Certificate of Service — Volunteer
WSO2 Certified Micro Integrator Practitioner V4
Google GDG Certificate
Become a Linear Algebra Master
A Deep Understanding of Deep Learning (with Python)
Python
JavaScript
TypeScript
Swift
Java
C
C++
R
SQL
PHP
HTML
CSS
Shell
PyTorch
TensorFlow
Hugging Face
LangChain
LlamaIndex
scikit-learn
PyTorch Geometric
Gradio
Streamlit
UMAP
LLM OptimisationSLM Fine-tuningRAG
Prompt EngineeringTransformersAttention MechanismsGANs
VAEsDiffusion ModelsGNNs
On-Device AI
Deepfake DetectionQuantisationPruningDistillationEfficientNetVision Transformer
MediaPipe
OpenCV
TesseractCLIPDocument AnalysisOCRSwiftSwiftUI
Core MLiOS DevelopmentVision FrameworkAVFoundationCore DataMetal MPSOn-Device Inference
React.js
Next.js
FastAPI
Node.js
Express.js
TailwindCSS
ViteReact Native
WebSocketREST APIs
AWS
Azure OpenAI
Azure AD
Google Cloud
Docker
Vercel
WSO2
CI/CD
Git
GitLab
Serverless
PostgreSQL
MySQL
MongoDB
Redis
Firebase
Supabase
Pandas
NumPy
MatplotlibFAISSSynthetic DataI write AI from Scratch, a blog series that derives machine learning models from the maths to the training loop. Smaller experiments and code are on GitHub.