Rohan Sardar
AI · ML · Data Science · Data Engineering
B.Tech CSE student specializing in AI & ML at Techno India University. Developer of multi-agent orchestration reasoning graphs, asynchronous retrieval pipelines, and deep translation models. Focused on building high-efficiency, production-ready AI services.
Work Experience
Built and optimized solutions for ML benchmarking tasks involving tabular learning, NLP, RAG, forecasting, and prompt engineering to surpass baseline AI agent performance.
Published technical articles on their official website.
Projects
Finance Verifier Agent
Built a multi-agent financial analysis system using LangGraph and Gemini 2.5 Flash to verify financial claims, fetch stock data, and generate structured analyst reports. Implemented real-time streaming with FastAPI WebSockets and dynamic PDF generation for report downloads. Achieved 88.9% routing accuracy and 77.8% tool execution accuracy with an average system latency of 8.56s.
View RepositoryCustom Asynchronous RAG
Developed a custom RAG pipeline with document ingestion, chunking, FAISS-based retrieval, and local FlashRank reranking to improve retrieval quality without additional API costs. Designed an asynchronous architecture with isolated vector stores and conversational memory for long-running queries.
View RepositoryTransformer from Scratch
Implemented a sequence-to-sequence Transformer architecture from scratch based on the “Attention Is All You Need” paper, including custom encoder, decoder, and attention mechanisms. Built an end-to-end training pipeline for English-Bengali machine translation.
View RepositoryMultimodal Semantic Search
Developed a semantic image search engine using OpenAI CLIP embeddings and Qdrant vector database for text-to-image and image-to-image retrieval. Built scalable retrieval APIs and interactive search workflows for contextual similarity matching.
View RepositoryToxic Comment Classifier
Built a real-time multi-label toxicity detection API using TF-IDF vectorization and Logistic Regression models trained on the Jigsaw dataset. Containerized the modular inference pipeline for scalable deployment.
View RepositoryPublications & Articles
Visualizing ResNet18 Attention with Class Activation Mapping (CAM) in PyTorch
Dec 2025Class Activation Maps (CAM) identify input image regions driving CNN class decisions. CAMs highlight spatial locations in feature maps most influential for predictions, allowing interpretable model visualizations.
Read on Medium →Visualizing BERT Word Embeddings with NetworkX and Plotly using PyTorch
Dec 2025Explores extracting hidden states from pre-trained BERT, computing cosine similarity matrices, and plotting interactive semantic network graphs via NetworkX and Plotly.
Read on Medium →Stratified K-Fold: Resampling Technique for Variance Reduction in Imbalanced Data
Mar 2026Demonstrates utilizing Stratified K-Fold validation to preserve target distribution ratios across folds, providing stable model generalization metrics on imbalanced datasets.
Read on Medium →Technical Skills
Languages
Python
C / C++
R
SQL
ML & Deep Learning
PyTorch
TensorFlow
Scikit-Learn
XGBoost
Transformers
Data & Big Data
NumPy
Pandas
Matplotlib
Statsmodels
PySpark
Databricks
Apache Airflow
MySQL
Systems & Deploy
FastAPI
LangChain
LangGraph
Vector Databases
Docker
Git
Google Cloud
Education
Techno India University
Kolkata, WB
Bachelor of Technology in Computer Science Engineering
Specialization in Artificial Intelligence & Machine Learning
Achievements
-
🏆Kaggle Notebooks Expert
Ranked in the top 1% globally in the Notebooks category.
-
🏅Finalist at Neurathon 26
Hackathon organized by ML Club of NIT Silchar.