Skip to content
View bk1210's full-sized avatar
🤍
🤍

Highlights

  • Pro

Block or report bk1210

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
bk1210/README.md


Portfolio Resume LinkedIn HuggingFace Gmail Instagram



$ whoami

Int. M.Sc. Data Science student at Amrita Vishwa Vidyapeetham, Coimbatore (2022–2027) working at the intersection of deep learning, sports analytics, and data analytics. I build systems that don't just predict — they explain.

  • 🏛️ Degree — Int. M.Sc. Data Science
  • 📍 Location — Coimbatore, Tamil Nadu 🇮🇳
  • 🔭 Focus — NLP · Data Analytics · Sports AI
  • 🚀 3 Live Deployed Apps on HuggingFace Spaces
  • 📦 9 Public GitHub Repositories
  • 💬 Languages — Tamil · English · Malayalam


$ ls projects/


01 · Graph DL · Cricket

Spatio-temporal graph learning for IPL T20 match outcome prediction. GAT player-interaction graph (652 players, 7,334 edges) + BiLSTM + Cross-Attention Transformer. Ball-by-ball win probability across 225,954 deliveries, 950 matches.

AUC-ROC 0.7378 · Val Acc 69.33% · SHAP-identified top predictors · Optuna-tuned

PyTorch PyG SHAP Optuna


02 · NLP · Transformers

Multi-task RoBERTa with novel Sentiment Incongruence Auto-Labeler. Dataset-agnostic sarcasm labels from semantic mismatch between surface sentiment and underlying emotion. LIME token-level rationale extraction on every prediction.

F1-Macro 0.977 · AUC 0.997 · 4,293 test samples · LIME explainability

RoBERTa HuggingFace Focal Loss LIME


03 · IoT · Football

Classifies football players into Attacker / Midfielder / Defender from PAMAP2 IoT wearable sensor data. Activity-to-role mapping pipeline. M3 TCN + Transformer architecture with SHAP explainability.

Accuracy 99.24% · LOSO 98.89% ± 0.42% · PAMAP2 dataset

TCN SHAP PyTorch


04 · IoT · ML

Three-class fatigue prediction (LOW/MEDIUM/HIGH) from 2.87M raw IoT sensor readings across 9 subjects. Novel Karvonen heart rate labeling scheme. 88 statistical features per sliding window. Coach substitution-alert dashboard.

Accuracy 97.87% · F1-Macro 79.93% · LOSO 97.96% ± 2.57% · Beats PAMAP2 benchmark ~88%

scikit-learn SMOTE TensorFlow


05 · RAG · LLM   🚀 Live Demo

Factually grounded UEFA Champions League match summaries. Self-curated UCL-2025 dataset (189 matches × 142 cols) from uefa.com. LLaMA 3.1 via Groq API for hallucination-free generation. SHAP identifies goal difference as strongest predictor.

Cosine Similarity 0.903 vs 0.373 baseline · 189 UCL matches · SHAP mean |SHAP| = 1.000

LLaMA FAISS Groq Streamlit


06 · ML · Sports Analytics   🚀 Live Demo

Multi-league prediction across 5 European leagues. 33 features from 25,979 historical matches — ELO ratings, rolling form, betting market probabilities, H2H stats. 3-layer anti-leakage architecture. Predicts outcomes for 259 teams.

Acc 0.55 · xPts MAE 0.91 · +0.073 F1-Macro over ELO baseline · 5 leagues

Gradient Boosting ELO SHAP SQLite


07 · Reinforcement Learning

AI goalkeeper trained via PPO in a 2D Pygame football simulation. Rule-based threat assessment + shot prediction + state machine (IDLE → TRACKING → READY → DIVING → RECOVERING). Improves with every game.

PPO Pygame PyTorch


08 · Agentic AI   🚀 Live Demo

Agentic AI assistant for college students using Observe → Think → Act loop. LLaMA 3.3-70B via Groq API. Finds curated resources and generates personalized exam study plans with live progress tracking.

LLaMA Groq Streamlit Tool Calling


09 · NLP · Web App

AI-powered web platform for students with ADHD, dyslexia & learning disabilities. T5 PDF summarization, spaCy quiz generation, TTS, multilingual translation across 7 Indian languages. Full-stack with Django + PostgreSQL.

IEEE Pegasus 3.0 Finalist ✅

Django T5 spaCy PostgreSQL



$ cat skills.json


PyTorch HuggingFace scikit-learn Python RAG SHAP GAT Power BI Streamlit PostgreSQL Kaggle LaTeX

$ github --stats







"The best moments are the ones you never planned."

— Cristiano Ronaldo




Open to research collaborations, internships, and exciting projects in AI, sports analytics, and deep learning.

🌐 Portfolio  ·  📄 Resume  ·  📬 bharathkesav1275@gmail.com

Snake animation

Popular repositories Loading

  1. Football-Goalkeeper-Training-Using-Reinforcement-Learning Football-Goalkeeper-Training-Using-Reinforcement-Learning Public

    AI goalkeeper learns to make saves through PPO Reinforcement Learning in a 2D Pygame football simulation. Rule-based threat assessment + shot prediction + state machine behavior (IDLE→TRACKING→READ…

    Jupyter Notebook

  2. Smart-Resource-Finder-Agent Smart-Resource-Finder-Agent Public

    Agentic AI study assistant for college students. Finds curated learning resources (videos, docs, tutorials, papers) for any academic topic and generates personalized exam study plans with progress …

    Python

  3. Detecting-Sarcasm-as-Sentiment-Incongruence Detecting-Sarcasm-as-Sentiment-Incongruence Public

    Multi-task RoBERTa for sarcasm detection via Sentiment Incongruence Auto-Labeling. Focal Loss + WeightedRandomSampler + LIME explainability. F1-Macro 0.977, AUC 0.997. Zero-shot cross-domain eval o…

    Jupyter Notebook

  4. CricketGraph-DL-IPL-Match-Outcome-Prediction-Player-Impact-Analysis CricketGraph-DL-IPL-Match-Outcome-Prediction-Player-Impact-Analysis Public

    Spatio-temporal graph deep learning for IPL T20 match outcome prediction. GAT player-interaction graph + BiLSTM + cross-attention Transformer. Ball-by-ball win probability, run forecasting & Player…

    Jupyter Notebook

  5. Football-Player-Fatigue-Prediction-Wearable-IoT-Sensors-ML Football-Player-Fatigue-Prediction-Wearable-IoT-Sensors-ML Public

    Three-class football player fatigue prediction from PAMAP2 wearable IoT data. Karvonen heart rate labeling, SMOTE balancing, LOSO cross-validation, personalized Random Forest. 97.96% LOSO accuracy …

    Jupyter Notebook

  6. FootballRole-DL FootballRole-DL Public

    Classifies football players into Attacker, Midfielder, Defender roles from PAMAP2 IoT wearable data. LSTM, BiLSTM, and TCN-Transformer architectures. 99.24% accuracy, LOSO 98.89%±0.42%. SHAP sensor…

    Jupyter Notebook