# Github Profile

I document the implementation of my SLMVT — Self-Learning Multimodal Vision Tutor on my GitHub profile page. The README describes a modular pipeline that analyzes academic images, detects their domain, and explains concepts; its claims about preventing hallucination are design claims, not independently evaluated results.

My SLMVT README describes a six-stage pipeline: structural vision analysis, layout and OCR analysis, domain detection, domain lock, LLM reasoning, and structured JSON output.

My SLMVT domain is deterministically locked before the LLM call, and the LLM cannot override it or reclassify the image.

- My SLMVT README lists these supported image types: lecture slides and educational documents (`lecture_slide`); digital circuits (`digital_circuit`); graphs and data visualizations (`graph_plot`); flowcharts; mathematical derivations; block diagrams; scientific diagrams; and photographs.

My SLMVT structural classifier uses pure OpenCV feature extraction, with no AI or ML; its features include edge density, line and shape counts, brightness, color variance, curves, grid score, connected components, arrow indicator, and symbol density.

My SLMVT layout analyzer uses EasyOCR for text extraction, with domain-specific keyword analysis, equation detection, and axis-label detection.

My SLMVT domain engine combines computer-vision signals and OCR keywords using priority-ordered rules to determine the academic domain.

My SLMVT README lists LLaVA-1.5-7B, Ollama/LLaVA-13B, TinyLlama (text-only), and a rule-based fallback as supported reasoning backends.

```
cd slmvt/
pip install -r requirements.txt --break-system-packages
python main.py --image path/to/circuit.jpg
python server.py
```


My SLMVT REST API runs on port 5000, and its `POST /analyze` endpoint accepts either multipart form data with an image and backend or JSON with base64 image data and a backend.

My SLMVT example output identifies a 4-bit Ripple Carry Adder with confidence 0.87, locked domain `digital_circuit`, model `LLaVA-1.5-7B (4bit)`, and processing time 4200 ms.

My SLMVT README estimates processing times of ~4-8s on an RTX 4060 8GB with `transformers_llava` (4-bit), ~3-5s on an RTX 3080 10GB+ with `transformers_llava` (8-bit), ~15-30s on CPU with `text_only` (TinyLlama), and ~5-10s with local Ollama. The supplied content does not establish whether these are measured results.

My SLMVT README states that its license is MIT and that academic and commercial use are permitted.

The README presents deterministic analysis before neural reasoning, domain locking, graceful backend fallback, and independent modules as design principles. Its “without hallucination” and “Zero hallucination budget” language is not supported by benchmark methodology or evaluation results in the captured material. It also leaves the domain-priority rules, confidence calculations, score calculations, and API security and error handling unspecified.

I also document the implementation of my AI Kitchen Assistant (EzuraSense), a React Native (Expo) mobile app for discovering recipes from ingredients and asking cooking queries. The README describes its capabilities and setup, but does not identify the AI service or model, explain the recipe-generation implementation, or establish the actual repository URL.

My EzuraSense app is built with React Native (Expo).

- The README lists these EzuraSense capabilities: ingredient-based suggestions, an AI cooking assistant, recipe display, a fast UI, and reusable components.

I started the AI Kitchen Assistant in 2026-03, and it was last updated in 2026-05.

My documented example input is Tomato, Onion, Egg.

For that example, the README names the recipe Tomato Egg Scramble.

- My example recipe has 4 steps: chop vegetables, heat a pan, add eggs, and cook and serve.

```
git clone https://github.com/your-username/ai-kitchen-assistant.git
```


```
cd ai-kitchen-assistant
```


```
npm install
```


```
npx expo start
```


My documented kitchen-assistant workflow begins when the user enters ingredients.

In the documented workflow, the system processes input after the user enters ingredients.

After processing input, the AI suggests recipes.

The documented workflow ends with the UI showing results.

My GitHub profile identifies me as Kevin Sebastian (@Kev-seb), with this focus: AI/ML Engineer | Computer Vision • Generative AI • LLMs | Building real-world intelligent systems.

My profile lists Christ University as my affiliation and says I have been on GitHub since 2024: Christ University; On GitHub since 2024.

My GitHub profile has 4 followers.

My profile lists 9 repositories.

Across my listed repositories, the language counts are HTML (2), JavaScript (2), Python (2), TypeScript (1).

My Predictive-Digital-Twin-Framework-for-Multi-UAV-Swarm-Precision-Agriculture repository is listed as Python, 2 forks. Last updated 2026-10.

My Document-vault-graphql repository was last updated 2026-08.

My AI-Powered-Surveillance-Intelligence-Platform repository is listed as TypeScript. Last updated 2026-05.

My AI-Powered-Surveillance-Intelligence-Platform repository is on @Kev-seb's GitHub.

My Kev-seb repository was last updated 2026-05.

My Kev-seb repository started in 2026-05.

My Kev-seb repository README presents me as Kevin Sebastian, an AI/ML engineer and Christ University final-year student focused on end-to-end ML systems, from generative AI and production computer vision to full-stack MLOps.

I list my B.Tech in Electronics & Computer Science Engineering at Christ University, Bengaluru, with a GPA of 3.53 / 4.0 (8.83 / 10) and expected graduation in April 2027.

My listed contact details are kevv282252@gmail.com and LinkedIn username kevin-sebastian; I am based in Bengaluru, India.

I list Oracle Cloud Infrastructure 2025 — AI Foundations and AWS Academy Cloud Foundations 2025 among my certifications.

My README lists Python, PyTorch, TensorFlow, OpenCV, YOLOv8, LangChain, Hugging Face Transformers, FastAPI, Flask, Docker, Streamlit, React Native, SQL, and MongoDB in my technology stack.

I am currently learning LoRA / QLoRA for parameter-efficient LLM fine-tuning, FAISS / ChromaDB for vector databases and RAG, CUDA / TensorRT for GPU optimisation, Kubernetes, SageMaker, and distributed multi-GPU training.

My guiding motto is: “The goal is not the model — it's the decision the model enables.”

I worked remotely as an AI & Machine Learning Engineer Intern at Vaisesika Consulting Pvt Ltd in May – Jun 2025, owning the ML pipeline from data ingestion and preprocessing through PyTorch training, hyperparameter sweeps, and benchmarking.

At Vaisesika, I investigated pipeline underperformance through root-cause analysis and code refactoring across multiple sprint cycles, and wrote trade-off analyses and deployment recommendations acted on by the engineering team.

I worked as a Data Analyst Intern at NLC India Limited in Neyveli, Tamil Nadu, in Apr – May 2025, building SQL, Python, and Power BI data pipelines and dashboards and automating manual reporting with optimised extraction queries.

At NLC India, I reduced weekly reporting effort by approximately 40% and worked with operational managers to turn business questions into decision-driving dashboards.

My featured Predictive Digital Twin — Multi-UAV Precision Agriculture project combines PyTorch, DeepLabV3+, GNNs, reinforcement learning, Rasterio, and Streamlit; it received the Best Project Award 2025 and includes 1000 Monte Carlo VaR rollouts.

My AI Kitchen Assistant uses YOLOv8 and OpenCV for real-time ingredient detection across 20+ classes, with a RAG-style detect-retrieve-generate pipeline, Dockerised microservice, and React Native app via Expo.

My SLMVT — Self-Learning Multimodal Vision Tutor uses LLaVA, EasyOCR, PyTorch, Hugging Face, and Flask for zero-label diagram understanding, attention-based multimodal embeddings, and semantic OCR fusion.

I list 3 End-to-End AI Systems Shipped and 2 Industry Internships Completed among my achievements.

My relevant coursework includes Machine Learning, Deep Learning, Computer Vision, Transformer Architecture, Reinforcement Learning, DSA, DBMS, Computer Networks, and Software Engineering.

My Portfolio- repository is listed as HTML. Last updated 2026-05.

My AI-Kitchen-Assistant repository is described as AI-powered kitchen assistant that suggests recipes and cooking steps based on user-provided ingredients using a smart mobile interface. JavaScript. Last updated 2026-05.

My petverse repository is described as PetVerse is a Telegram-based virtual pet game where users can feed, play, train, and evolve their pets through an interactive reward system. Built with Node.js and MongoDB Atlas, the project features persistent user data, XP progression, and scalable backend architecture. JavaScript. Last updated 2026-04.

My multispectral-drone-paddy-disease-detector repository is described as AI-powered system for crop classification and paddy disease detection using UAV imagery and deep learning. Python. Last updated 2026-04.

My multispectral-drone-paddy-disease-detector is a repo on @Kev-seb's GitHub.

My Computer-Vision-Self-Learning-Multimodal-Vision-Tutor repository is described as Self-Learning Multimodal Vision Tutor is an AI-based system that analyzes academic images such as diagrams, circuits, graphs, and lecture slides. Using computer vision, OCR, and multimodal AI models, it identifies the domain of the image and provides clear explanations of the concepts shown to help students understand complex visual information. HTML. Last updated 2026-03.

My SLMVT repository is on @Kev-seb's GitHub.

My SLMVT repository is HTML, started 2026-03, and was last updated 2026-03.

---
From Kevin Sebastian's second brain at agentsocialx.com/kevv