SDE · AI/ML Engineer · Researcher · ECE @ NIT Patna

ShreyanshGupta

I work across |

ECE @ NIT Patna · Building ML systems and full-stack products · Published at IEEE MIGARS 2026 and MIND 2025 · Amazon MLSS ’25 · Codeforces Specialist

2Accepted Papers
0+DSA Solved
8.38CGPA / 10
Top 3KAmazon MLSS '25

Open to SDE / SWE and AI-ML roles — 2026 internships, 2027 new grad · Patna, India · Graduating June 2027

01About

Profile

Electronics and Communication undergraduate at NIT Patna, working across machine learning research and full-stack engineering.

My research combines graph structure with language models. At the BioML Lab I co-authored a MIND 2025 paper fusing BERT text embeddings with Node2Vec citation-graph embeddings, reaching 90.56% accuracy on the ACL Citation Sentiment Corpus.

For an ISRO-sponsored minor project I built Transformer and Mamba–Transformer architectures to classify winter wheat growth stages from Sentinel-1 SAR and Sentinel-2 optical time series, benchmarking eight models across 204 fields in Bihar.

I also build and ship. Paridhi is a civic-infrastructure PWA with a LangGraph discovery agent and zero-token semantic search — it took a six-person team I led to the India Innovates 2026 national finals. MiniStream is a full-stack MERN streaming app. Both are deployed and public.

02Work

Experience

Research Intern

Field-Boundary Delineation from Multi-Temporal Sentinel-2 Imagery

IIT Roorkee — Agricultural remote sensing

  • Built a delineation pipeline for smallholder field boundaries from multi-temporal Sentinel-2 imagery — Google Earth Engine ingestion, cloud masking and label preparation through to training and evaluation.
  • Benchmarked segmentation architectures and temporal configurations under a single evaluation protocol.
ML Trainee

Amazon ML Summer School 2025

Amazon — Selected among the top 3,000 of 60,000+ applicants

  • Selected for Amazon's ML Summer School — top 3,000 of 60,000+ applicants nationally.
  • Trained by Amazon Applied Scientists across Deep Learning, Generative AI, Reinforcement Learning, Causal Inference and LLMs, with a focus on implementation, experimentation and evaluation workflows.
Student Researcher

Citation Sentiment Analysis — MIND 2025 Paper

BioML Lab, NIT Patna — Under Prof. Akshay Deepak · NIT Patna

  • Co-authored Hybrid Graph–Language Embedding Framework for Citation-Aware Sentiment Analysis, accepted at MIND 2025.
  • Fused BERT text embeddings with Node2Vec citation-graph structural embeddings, processed through a Keras GCNN and PyTorch Geometric GCNConv/GATConv to classify citation sentiment as positive, neutral or negative.
  • Reached 90.56% accuracy on the ACL Citation Sentiment Corpus (8,736 instances), beating text-only and TF-IDF baselines across 8 classical classifiers including RF, SVM, AdaBoost and Extra Trees.
Open Source

GSSoC 2024 — AI/ML Contributor

GirlScript Summer of Code

  • Improved terrain recognition models via transfer learning → +7% accuracy.
03Publications

Research

Peer-reviewed work and ongoing studies.

  1. 01
    MIND 2025✓ Accepted

    Hybrid Graph–Language Embedding Framework for Citation-Aware Sentiment Analysis

    Anukul Kapoor, Shreyansh Gupta, Akshay Deepak · BioML Lab, NIT Patna

    Three-class citation sentiment (positive / objective / negative) on the ACL Citation Sentiment Corpus — 8,736 annotated instances over a 2,992-node directed citation graph. Fuses a Node2Vec structural embedding with a PCA-reduced BERT [CLS] sentence vector into a 262-dimensional node representation, then passes it through a Keras GCNN with GraphSAGE-style aggregators and PyTorch Geometric GCN / GAT / APPNP variants. Structure plus text beats text-only baselines across LR, RF, SVM, AdaBoost and ExtraTrees classifiers.

    Graph Neural Networks · BERT · Node2Vec · PyTorch Geometric · Citation Analysis

    Repository →
  2. 02
    IEEE MIGARS 2026✓ Accepted

    Crop Phenology Classification: A Multi-Paradigm Benchmark from Traditional Machine Learning to Hybrid Mamba–Transformer Networks

    Shreyansh Gupta, Ashutosh Panda, Shashank Pandey, Gautam Nath, Ajay Kumar Maurya · RAC-S, ISRO — under Prof. Ajay Kumar Maurya

    Winter wheat growth-stage classification from Sentinel-1 SAR and Sentinel-2 optical time series over 204 agricultural fields in Bihta, Bihar. Introduces two architectures — a Transformer with tripartite temporal embeddings, and a bidirectional Mamba–Transformer hybrid — and benchmarks eight models across three modality configurations, reaching a macro-F1 of 0.7765 through multimodal temporal fusion without temporal resampling.

    Remote Sensing · Mamba · Transformer · Sentinel-1 SAR · Sentinel-2 · Time Series

    Repository →
04Selected Work

Projects

01TypeScriptIndia Innovates 2026 Finalist

Paridhi — Civic Infrastructure Explorer

A mobile-first PWA for civic infrastructure discovery and community issue reporting. A LangGraph agent crawls OpenStreetMap and local news in parallel, extracts structured project facts via LLM, geocodes with Nominatim, then validates every result through a deterministic non-AI check. Adds zero-token semantic search over in-process ChromaDB embeddings, gravity-model + 2SFCA civic gap heatmaps, TSP-based walk planning and O(1) grid-hash geofencing.

40 unit tests
GitHub Actions CI
Zero-token search

React · TypeScript · LangGraph · ChromaDB · Gemini · Firebase · Leaflet

02MERNLive

MiniStream — Full-Stack Streaming App

Full-stack MERN streaming application with JWT authentication stored in httpOnly cookies (bcrypt-hashed), protected REST routes, TMDB-powered movie and TV browsing, trailer playback and per-user search history.

MongoDB · Express · React · Node.js · JWT · Zustand · Tailwind · TMDB API

03PyTorchISRO-Sponsored

Crop Phenology Classification

Minor project sponsored by RAC-S, ISRO under Prof. Ajay Kumar Maurya. Two architectures for winter wheat growth-stage classification from Sentinel-1 SAR and Sentinel-2 optical time series: a Transformer with tripartite temporal embeddings, and a bidirectional Mamba–Transformer hybrid.

macro-F1 0.7765
204 fields
8 models benchmarked

PyTorch · Mamba · Transformer · Sentinel-1 SAR · Sentinel-2 · Time Series

04ResearchPublished · MIND 2025

Citation Sentiment Analysis

Citation counts tell you how often a paper is cited, not how. This classifies each citation as positive, objective or negative by fusing two views: the citation sentence encoded with BERT, and the paper’s position in the citation graph encoded with Node2Vec.

8,736 citations
2,992-node graph
262-d fusion

PyTorch Geometric · BERT · Node2Vec · GCN · GAT · TensorFlow

05TypeScriptWalmart Sparkathon 2025

WalmartGenie — AI Shopping Assistant

GPT + Gemini APIs surface hidden-gem products, trending picks and best deals in seconds. Features virtual AR try-on for fashion items and a Firebase real-time backend.

React Native · OpenAI · Gemini · Firebase · AR Try-On

06PyTorchFeatured

FoodVision: Image Classification

EfficientNet-B2 (86.3%) + Vision Transformer (90.4%) on Food-101. Deployed as Gradio app with 0.15s inference.

ViT: 90.4%
EffNet: 86.3%
~0.15s

PyTorch · ViT · EfficientNet · Gradio

07TensorFlow

Violence Detection System

Fine-tuned MobileNetV3 achieving 89.4% accuracy with frame sampling + Streamlit UI for video analysis.

89.4% Accuracy
Kaggle: 89%

TensorFlow · MobileNetV3 · OpenCV · Streamlit

08C++

Hex Game — AI & Manual

11×11 Hex board game with Monte Carlo Tree Search (MCTS) AI opponent + DFS pathfinding.

C++ · MCTS · DFS · Game AI

09JupyterGSSoC 2024 · Open Source

Terrain Recognition (CNN)

Lightweight explainable CNN — improved +7% accuracy via transfer learning during GSSoC 2024.

CNN · XAI · Transfer Learning · Open Source

05Tooling

Tech Stack

Generative AI

  • LangGraph
  • ChromaDB
  • Retrieval-Augmented Generation
  • Gemini API
  • OpenAI API
  • Vector Search

Languages

  • Python
  • C++
  • C
  • TypeScript
  • JavaScript
  • MATLAB
  • LaTeX

ML Frameworks

  • PyTorch
  • PyTorch Geometric
  • TensorFlow / Keras
  • Scikit-learn
  • Hugging Face
  • OpenCV

Modelling

  • Transformers
  • Mamba / State Space Models
  • Graph Neural Networks
  • BERT
  • Node2Vec
  • Time-Series Modelling

Tools & Platforms

  • Git / GitHub
  • Google Earth Engine
  • AWS
  • Firebase
  • Vercel / Netlify
  • Postman

Web & Backend

  • React
  • React Native
  • Node.js / Express
  • MongoDB
  • MySQL
  • REST APIs

Also

Pandas · NumPy · SciPy · Matplotlib · Seaborn · TensorBoard · Gradio · Streamlit · Flask · FastAPI · REST APIs · MySQL · Rasterio · GeoPandas · MATLAB · LaTeX · Tailwind · Vite · Astro

06Record

Achievements

  • Amazon ML Challenge 2025

    A separate Amazon competition with its own problem statement. Ranked under 800 of 80,000+ participants, peak live rank 43.

  • Amazon ML Summer School 2025

    Selected among the top 3,000 of 60,000+ applicants for Amazon's national machine learning training programme.

  • ATF Fellowship

    Selected among the top 450 of 20,000+ applicants for Stage 2.

  • Competitive Programming

    450+ problems solved on LeetCode (rating 1718) · CodeChef 2-Star · Codeforces Specialist.

  • Competitive Examinations

    GATE DA 2026 AIR 5,262 · JEE Advanced 2023 AIR 17,222 · JEE Main AIR 19,582.

  • Three National Finals

    Qualified for India Innovates 2026, ScrollHacks and Hack This Fall — best performance Top 16 at ScrollHacks.

  • HackSlash AI/ML Group

    Member of Team Gray Interface, HackSlash AI/ML Group at NIT Patna since 2023.

  • ISRO-Sponsored Research

    Minor project sponsored by RAC-S, ISRO on crop phenology from Sentinel-1 and Sentinel-2 time series.

  • Two Accepted Papers

    MIND 2025 on citation-aware sentiment analysis, and IEEE MIGARS 2026 on crop phenology classification.