Hello, I'm Aritro Shome
AI Engineer & Programmer
I build intelligent systems by translating complex mathematical theories into elegant, functional code. I love building efficient tools which solve problems optimally with maximum results.
The Beauty of the Equation
For me, AI and Machine Learning aren't just about black-box abstractions; they are the living application of mathematics. My passion lies in building key algorithms from scratch to reinforce conceptual understanding—from the calculus that drives optimization, to the linear algebra that defines vector spaces, and the probability that governs uncertainty. I strive to build models that are not only powerful but also mathematically sound and implemented from first principles. I also love learning about the technological diaspora which enables me to tackle novel problems efficiently and produce optimal results with reliable accuracy and precision. My BLEU and ROUGE would be my peers and educators who've seen me become whatever I am today and I am grateful to them for their valuable contributions.
Education
Indian Institute of Engineering Science and Technology, Shibpur
Bachelor of Technology, Information Technology
CGPA: 8.94
Aug 2024 - May 2028
M. P. Birla Foundation Higher Secondary School
Secondary & Higher Secondary Schooling
ISC (Class 12): 95.3%
ICSE (Class 10): 98.3%
2010 - 2024
Inspire Scholarship
Selected by Government of India
May 2024
Experience
AI Engineer Intern
Sarvam AI
Aug 2025 – Dec 2025
- Spearheaded the evaluation framework for an NVIDIA-provided LipSync Model, developing both quantitative and qualitative metrics and automating the quantitative pipeline with Python and data visualization to benchmark performance.
- Diagnosed and resolved a critical bug in the LipSync model's output pipeline that caused unpredictable video file size changes (~500% variance), ensuring stable and reliable video generation.
- Redesigned the dubbing pipeline's translation prompt architecture by identifying and modularizing redundancies, contradictions, and hard-coded elements in the existing prompt library.
- Implemented a YAML-based, block-wise structure for prompt composition to improve maintainability and proposed changes, significantly streamlining future prompt modifications across the system.
- Executed A/B testing on the newly implemented modular prompt architecture and discovered that an unformatted "open-free" prompt achieved superior translation quality, leading to an optimized and simplified final deployment.
Featured Work
Where theory meets application.
AlponaGen - Procedural Generative Art Engine
Designed and engineered a modular Python engine to procedurally generate Alpona, a traditional Bengali folk art, using a fractal-based mathematical layering algorithm to produce intricate, authentic Alpona images.
- Generated and published a novel, high-quality synthetic dataset of 13,000 images (1024x1024) on Kaggle to facilitate research in culturally relevant generative AI and pattern recognition tasks.
- Implemented the core `style_alpona` algorithm which utilizes concentric layers and probabilistic style selection to control layer density, pattern, and aesthetic composition.
- Established a modular, object-oriented framework with a `Style Registration System` to decouple the generation engine from pattern drawing functions, ensuring extensibility and robust code maintenance.
Thomas Raw - Agentic AI Based Travel Planner
Developed an AI-powered travel assistant using Google Gemini and LangChain, designed to plan, budget, and personalize travel itineraries. Tech stack: Python (Flask), LangChain, Google Gemini API, HTML/CSS/JS.
- Implemented multiple modules including Mr. Thomas Chats (conversational agent), Budget Buddy (cost estimator), and Where Is It? (photo-based location identifier).
- Integrated export functionality through the Insight Centre, enabling users to save and share their generated trip plans.
Predicting the 2025 IPL Winner
Explored how data science could be applied to sports prediction and fan engagement. Built models including Decision Trees, Random Forests, MLP Classifiers, and Monte Carlo simulations.
- Performed comprehensive EDA and model evaluation to maximize prediction accuracy.
- Won BrainDead 2025 Hackathon organized by the CST Department, IIEST Shibpur.
Abler (Winner - SXC eXabyte Hackathon)
Built a real-time accessibility assistant with object detection, depth estimation, and speech/gesture recognition to address accessibility gaps for the visually impaired using multimodal AI systems.
Certifications
Continuous learning and specialization.
My Toolkit
The languages, frameworks, and concepts I use daily.
Languages
- Java
- Python
- C/C++
- JavaScript, HTML/CSS
Frameworks
- Flask
- Numpy
- Pandas
- PyTorch
- Scikit-Learn
- Matplotlib
Developer Tools
- Git
