Hi, I'm Shubh Umre, a B.Tech undergraduate specializing in Artificial Intelligence & Machine Learning.
I build practical AI applications and explore the engineering behind them — from data preparation and model development to retrieval pipelines, intelligent interfaces and deployment.
My current focus is on:
|
Machine Learning |
LLM Applications |
Agent Workflows |
APPLIED AI · GENERATIVE AI · AGENTIC AI
The map below is rendered directly by GitHub as a Mermaid diagram, rather than relying on an externally embedded SVG.
flowchart LR
A["🧠 Applied AI / ML"] --> B["Data & Model Development"]
C["✦ Generative AI"] --> D["LLMs · Embeddings · RAG"]
E["⚡ Agentic AI"] --> F["Tools · Memory · Workflows"]
B --> G(("AI Applications"))
D --> G
F --> G
G --> H["Deployment & Evaluation"]
classDef applied fill:#083344,stroke:#22d3ee,color:#cffafe,stroke-width:2px
classDef gen fill:#2e1065,stroke:#c4b5fd,color:#ede9fe,stroke-width:2px
classDef agent fill:#3b1d0d,stroke:#fb923c,color:#ffedd5,stroke-width:2px
classDef core fill:#102b3c,stroke:#34d399,color:#d1fae5,stroke-width:2px
class A,B applied
class C,D gen
class E,F agent
class G,H core
A selection of my AI/ML applications and technical resources. Open each repository for implementation details, architecture and setup.
|
HYBRID RETRIEVAL · GENERATIVE AI
A document-grounded question-answering application. Upload PDF, TXT, DOCX or CSV files and ask questions in natural language. Combines Qwen3 embeddings and FAISS semantic search with BM25 keyword retrieval through LangChain's EnsembleRetriever.
|
COMPUTER VISION · BEHAVIORAL ANALYTICS
An experimental system that analyzes examination video for face movement, centroid-based motion, fidgeting signals, facial-expression signals and temporal patterns. These signals are combined into an experimental Anxiety Index (0–100).
The index is an experimental behavioral indicator, not a medical or psychological diagnosis. |
|
UNSUPERVISED ML · DATA INTELLIGENCE
Uses UIDAI demographic and biometric update datasets to study district-level and temporal patterns. The workflow includes data cleaning, monthly aggregation, feature engineering and Isolation Forest-based anomaly detection, with risk views in a Streamlit dashboard.
|
NATURAL LANGUAGE PROCESSING · CLASSIFICATION
A text-classification web app using NLTK preprocessing, TF-IDF feature extraction and a Multinomial Naive Bayes classifier. The project README documents the SMS Spam Collection dataset and a Streamlit prediction interface.
|
|
RECOMMENDER SYSTEMS · NLP
A content-based recommender that converts movie metadata into TF-IDF vectors and uses cosine similarity to find similar movies. The Streamlit interface also uses the TMDB API to retrieve movie posters.
|
VISUAL TECHNICAL RESOURCE · AI / ML
A collection of visual reference sheets covering Python, ML libraries, deep learning, computer vision, NLP, GenAI, RAG, embeddings, vector databases, LangChain, agents, LLM evaluation, fine-tuning, LoRA and MLOps.
|
Repository insights, language distribution and contribution activity are generated from GitHub data by GitHub Actions.
I'm interested in practical work and collaboration across Machine Learning, Generative AI, Agentic AI, Computer Vision and open-source software.
