┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ whoami │ └─────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ cat about.md │ └─────────────────────────────────────────────────────────┘
> B.Tech CSE (AI Specialization) @ Amrita Vishwa Vidyapeetham, Coimbatore — Class of 2027
> Deep in campus placements right now — RL, ML systems, and full-stack engineering
> Faculty-assigned research: Memory-Aware RL — chemotactic escape from information
traps, sitting at the intersection of reinforcement learning & statistical biophysics
(Kramers escape rate theory, non-Markovian dynamics, agent-based search strategies)
> Prefer minimal, correct code over clever abstractions — competitive programming
mindset carries into everything I build
> Fluent in going from "read the paper" to "reproduce the result" to "ship the demo"
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ ls stack/ --tree │ └─────────────────────────────────────────────────────────┘
Languages
ML / RL / Data
Full-Stack / Infra
Tools
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ ./featured_projects.sh │ └─────────────────────────────────────────────────────────┘
|
Custom Gymnasium environment — 2D point-robot, domain-randomized obstacle avoidance — trained with PPO (Stable-Baselines3). 88.3% success rate over 300 held-out episodes, generalizing across obstacle count. Reward-shaping and debugging fully documented.
|
Real-time perception-to-risk pipeline on the KITTI driving dataset — CUDA-accelerated monocular depth (MiDaS), YOLOv8 detection + SORT tracking, and Essential-Matrix pose estimation fused into a live collision-risk dashboard.
|
|
Distributed big-data pipeline detecting suspicious Ethereum wallets via unsupervised ML. Docker + Hadoop HDFS + Apache Spark (Scala) ingesting the ORBITAAL dataset with engineered wallet-level features.
|
Full-stack live monitoring dashboard. React frontend visualizing global attack telemetry (region/type/severity) from an Azure SQL backend, with a Flask + Isolation Forest anomaly engine.
|
|
Role-based IoT energy monitoring platform built end-to-end: Node.js/Express API, SQLite, and a React + TypeScript + Tailwind + Recharts frontend. Built to prove out a deployed, production-shaped full-stack app.
|
Simulates a CDN-style network with NetworkX + Plotly — generates subnetworks, distributes files, and models real-time congestion with congestion-aware shortest-path retrieval.
|
|
MILP-driven decision-support dashboard for e-commerce logistics — facility location, transportation planning, seller assignment, and sensitivity analysis on real Olist data.
|
Agentic assistant on FastAPI + Next.js using the Model Context Protocol to connect LLMs to local tools — semantic + episodic memory via MongoDB/ChromaDB, task orchestration via Celery/Redis.
|
|
Multimodal deep learning pipeline for atrial fibrillation detection from dual-channel ECG + PPG signals. Converts WFDB recordings into log-STFT spectrograms and classifies rhythm irregularities with a CNN + BiLSTM hybrid, including subject-level splitting and full ROC/PR/threshold evaluation.
|
End-to-end pipeline for 3D CFD flow simulation using Fourier Neural Operators, a 3D autoencoder, and Diffusion Transformers. Covers mesh-to-grid preprocessing, latent-space refinement, DDIM sampling, and physics-aware flow reconstruction.
|
|
Research-grade framework combining classical time series models (ARIMA/SARIMA/VAR/VECM/GARCH) with deep learning (LSTM/GRU/CNN-LSTM) and hybrid ensembles to detect early-warning signals and forecast climate tipping points from ERA5-style temperature and precipitation data.
|
RAG chatbot built with LangChain, Gemini, FAISS, and Streamlit. Supports runtime PDF/CSV uploads that update the vector knowledge base on the fly, plus multi-turn conversational memory — no restart needed to add documents or maintain context.
|
|
Conversational BI agent analyzing live monday.com Work Orders and Deals data via Groq-powered tool calling — data normalization, quality checks, pipeline insights, receivables analysis, and automated leadership reports.
|
Platform for commercial intelligence, AI-powered advertising, knowledge graphs, and workflow/lifecycle automation. FastAPI + React + PostgreSQL + Docker with JWT auth and Google Vertex AI Imagen for production-quality marketing creative generation.
|
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ ls system-design/ --practice │ └─────────────────────────────────────────────────────────┘
|
Low-level design in Java — book cataloging, member registration, and issue/return workflows using clean OOP and SOLID principles, with repository-based storage abstraction and custom domain exceptions.
|
High-level design for a scalable news feed system (Facebook/Instagram/Twitter-style) — hybrid push/pull feed generation, DB schema, API design, caching, sharding, and fault tolerance.
|
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ git log --stats --graph │ └─────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ cat contribution_grid.snake │ └─────────────────────────────────────────────────────────┘
┌─────────────────────────────────────────────────────────┐ │ guest@tsar0705:~$ echo $CURRENTLY_LEARNING │ └─────────────────────────────────────────────────────────┘
[research] Kramers escape-rate theory → non-Markovian (Grote-Hynes) extensions
[applied] RL for continuous control — PPO reward shaping, domain randomization
[systems] Parameter-efficient adaptation for large audio-language models
[practice] Backtracking, DP, and output-ordering edge cases (competitive programming)
"An agent that learns to escape a trap is still, underneath it all, just solving for the gradient."