I am an IT student at SAL College of Engineering (CGPA: 8.93) based in Ahmedabad. I enjoy building end-to-end AI and ML systems that go from idea and data to a real working product. If you want to talk about AI, ML or GenAI, my DMs are open.

Portfolio: okyashgajjar.vercel.app | LinkedIn: linkedin.com/in/okyashgajjar | Kaggle: kaggle.com/okyashgajjar | Hugging Face: huggingface.co/okyashgajjar

What I do: I build modular ML pipelines, fine-tune LLMs with LoRA, QLoRA and Unsloth, and ship GenAI agents with LangChain, LangGraph and n8n. What I make: AI career advisors, civic AI agents, medical chatbots, low-spec Linux dotfiles, and tooling for coding agents.

I work across the full ML lifecycle, not just one part of it:

Supervised learning (practical): Regression, classification and ensembles. I have organized 15+ algorithms by their math families in Unified-ML-Pipelines with parallel execution and multi-metric evaluation.

Unsupervised learning (practical): Clustering and dimensionality reduction. Example is bbc-topic-discovery using Word2Vec and K-Means on the BBC News dataset.

Reinforcement learning (theory): I have studied MDPs, policy and value iteration and explored them conceptually in EvoPolis, a simulation of an AI-driven society.

Transfer learning (practical): Using pretrained models for new tasks, like ResNet50V2 for pneumonia detection and transformers for NLP.

NLP and Transformers: I built a fake news detector with TF-IDF and Ridge classifier that reaches 99 percent accuracy in Fake-Real-News-Detector and the full-stack fake-real-news-classifier with FastAPI and React.

AI Fundamentals: Probability, optimization, evaluation and feature engineering. I am active on Kaggle and I published When More Data Hurts (IJCSPUB, ISSN 2250-1770) about noise and scaling limits in ML.

Fine-tuning: I use Unsloth with LoRA and QLoRA. On huggingface.co/okyashgajjar you can find gemma-4-medical-full (5B, Image-Text-to-Text), gemma-medical-lora, the dataset doctor-patient-conversation-v1 (1.59k rows) and the Space Gemma 4 Medical Chatbot.

GenAI Stack: LangChain, LangGraph and n8n. My main project here is CivicLensAI, 1st Prize at HackTheSpark 2026, an agentic civic system where you upload an image and the agent detects the issue and routes it to the right department with severity and confidence. Other work includes innocheck-enterprise, ScreenSage-AI and open_deep_research_engine_v1.

End-to-End MLOps and Agent Tooling: costwise-mcp - a local MCP server for coding agents that adds repo intelligence, cache-write optimization and session memory. It has 21 stars and 4 forks in Go, and it powers the repo-aware pipeline behind my end-to-end workflow from data prep to evaluation to deployment.

Core Languages and Frameworks:

Also comfortable with: AI full stack development

yg@arch
-------
OS: Arch Linux x86_64
Host: 81WQ (IdeaPad 3 15IGL05)
Kernel: Linux 7.1.9-arch1-2
Uptime: 8 hours, 35 mins (profile Uptime = age: 20 years, 10 months, 16 days)
Packages: 1655 (pacman), 6 (flatpak)
Shell: bash 5.3.15
WM: niri 26.04 (Wayland)
Theme: Adwaita [GTK2/3/4] | Icons: Adwaita | Font: Adwaita Sans 11pt | Cursor: elementary 24px
Terminal: opencode
CPU: Intel Celeron N4020 (2) @ 2.80 GHz
GPU: Intel UHD Graphics 600 @ 0.65 GHz [Integrated]
Memory: 7.31 GiB (3.17 GiB used / 43%)
Disk (/): 166.12 GiB ext4 (71.97 GiB used / 43%)
hyfetch preset: aroace3 | mode: rgb | backend: fastfetch | lightness: 0.81

Hobby - Community Contributions and OpenSource: Hyprland/Niri low-spec dotfiles (around 230 MB idle) - Low-Spec-Niri-Dotfiles (6 stars, Shell, matugen Material You) and low-sepecs-hyprland-dotfiles (13 stars, CSS, boots under 300 MB with niri/waybar/matugen) plus feather-niri (Shell) - open-source hobby builds. Tested on IdeaPad 3 15IGL05.

• Unified-ML-Pipelines - math-driven, family-based ML pipelines with 15+ algorithms and parallel, multi-metric evaluation.
• ASPIRELY - DE-Aspirely-Updated - AI career advisor with voice mock interviews, 1st Prize at SAL College among 300+ teams, started as Tic Tech Toe 2025 proof of concept.
• CivicLensAI - 1st Prize at HackTheSpark 2026, agentic civic system where you upload an image and the LLM detects the issue and decides department, severity and confidence. Jupyter Notebook.
• costwise-mcp - local MCP server for coding agents with repo intelligence and cache optimization, 21 stars, 4 forks, Go, my most starred project.
• Medical Fine-tunes at HF - gemma-4-medical-full and gemma-medical-lora plus doctor-patient-conversation-v1 - see the Medical Chatbot Space.
• Fake News Detection - Ridge and TF-IDF with 99 percent accuracy, full-stack version is fake-real-news-classifier with React and FastAPI.
• When-More-Data-Hurts - code for my published paper PDF on label and feature noise and scaling.
• More: bbc-topic-discovery (NLP), pneumonia-detection-resnet50v2, telecom-churn (80 percent Logistic Regression), Medicine-Recommandation, EvoPolis, ScreenSage-AI, AI-dev-assistant.

Stats: 42 public repos, 45 stars, 17 followers, 17 following - github.com/okyashgajjar (since 2022-02-28). The card at the top is refreshed automatically by GitHub Actions.

• Pursuing B.E. in IT at SAL College of Engineering - CGPA 8.93
• Active Kaggle contributor, focused on feature engineering and end-to-end workflows
• 1st Prize at HackTheSpark 2026 for CivicLensAI (agentic civic AI)
• 1st Prize at SAL Project and Poster Presentation among 300+ teams for ASPIRELY, built from Tic Tech Toe 2025 hackathon
• Published When More Data Hurts (IJCSPUB, ISSN 2250-1770) - PDF
• Certified in Machine Learning and Data Analysis with Python (FreeCodeCamp)
• Hugging Face - 2 models, 1 dataset, 1 Space (Gemma medical, 1.59k rows)