PRAKASH
SINGH
I build complete systems end-to-end — from training CNNs on 9,000-image datasets and deploying quantized models to edge hardware, to self-hosting my own cloud infrastructure with Docker, DDNS, and reverse proxies. I ship real solutions.
About
I'm a 2nd-year B.Tech student who builds complete systems, not demos.
My focus sits at the intersection of deep learning, edge deployment, and self-hosted infrastructure. I've trained CNN pipelines on 9,000+ image datasets, quantized models to INT8 for NXP edge hardware, and shipped Dockerized workflows running on my own home lab.
I prefer measurable wins — accuracy numbers, hackathon results, deployed systems — over surface-level prototypes. Currently expanding my home lab and benchmarking local LLMs.
Skills
The toolchain behind the systems I ship.
Languages
AI / Machine Learning
Agentic & Local AI
Generative AI
Infrastructure & Systems
Dev Tools & Web
Core CS
Selected Work
End-to-end systems shipped to real users and real hardware.
Production-ready deep learning pipeline for automated semiconductor wafer defect detection. Classifies 8 defect types with 86.6% accuracy.
- ▸Trained on 9,000+ image dataset with synthetic augmentation via Stable Diffusion
- ▸MobileNetV2 + Squeeze-and-Excitation (SE) Attention architecture
- ▸Focal Loss training for class imbalance handling
- ▸INT8 quantization for edge deployment on NXP hardware (TFLite)
Syncs pinned tabs, tab groups, and workspace organization between two Zen Browser instances across dual-boot machines using Google Drive.
Built under strict competition constraints at KnowledgeQuarry — focused on correctness, efficiency, and clean problem decomposition under time pressure.
Award-winning simulation visualization tool for interpreting complex datasets with focus on clarity and cognitive accessibility.
Highly optimized AI game bot featuring alpha-beta search engine with pruning and a reinforcement learning training pipeline.
OpenEnv environment for deterministic vendor onboarding and compliance review with Docker-based reproducibility.
Mocktasium
AI-Powered Mock Test Platform
Responsive web application with AI-driven test generation and evaluation. Intelligent scoring, analytics dashboard, and adaptive question generation.
Achievements
Hackathon trophies and competitive recognitions.
Built VisualizeSim, a simulation visualization system for interpreting complex datasets.
Developed QuarryFlow Crossing under strict constraints — correctness & efficiency focused.
Production semiconductor defect detection AI pipeline deployed to NXP edge hardware.
National recognition in cybersecurity competition.
Experience & Training
AI Model Development
- ▸Trained models on Google Colab & Kaggle GPU environments
- ▸Built dataset curation, augmentation, and benchmarking pipelines
- ▸Optimized models for edge deployment via INT8 quantization (TFLite)
Local AI & Self-Hosting
- ▸Built offline AI pipelines and LLM assistants (llama.cpp, Ollama, LM Studio)
- ▸Configured home servers, port forwarding, DDNS, and Docker hosting
- ▸Automated workflows using AI-CLI tools and shell scripting
// Certifications
Half of what I actually know came from breaking things at 2am, not from these. But the paper helps.
Get In Touch
Open to internships, collaborations, and interesting systems problems.
Drop a message about an idea, an opportunity, or a build you want help with. I read everything and reply quickly.