Case Studies for AI / LLM Engineering
Each project is written for interview and recruiter review: Problem, My Ownership, Tech Stack, and Measurable Output.
Problem: Open-source LLM training projects are often hard to reproduce end-to-end in constrained local environments.
My Ownership: Independently executed the full MiniMind pipeline from pretraining to SFT, LoRA fine-tuning, and evaluation/inference.
Tech Stack: Python, PyTorch, distributed/multi-stage training scripts, checkpoint resume, and inference deployment scripts.
Measurable Output: Completed 4 core pipeline stages (Pretrain, SFT, LoRA, Eval) with reproducible local runs and validated checkpoint continuation workflow.
Problem: Real-world LLM applications require stable deployment chains, not just model demos.
My Ownership: Completed local environment setup, training data preparation, model tuning configuration, API service startup, and chatbot platform connection.
Tech Stack: Python ecosystem, model service APIs, retrieval/chat orchestration, and deployment scripts.
Measurable Output: Ran a 5-step deployment chain from data to service response, completed end-to-end verification, and published deployment evidence video.
Problem: Voice-to-task workflows often break between speech recognition and downstream action execution.
My Ownership: Integrated speech recognition and email task processing into a usable productivity-oriented AI prototype.
Tech Stack: Speech pipeline, workflow orchestration, and LLM-assisted task flow integration.
Measurable Output: Delivered a complete runnable loop: voice input -> transcript parsing -> email workflow execution.
Problem: Geometry education tools often lack progressive challenge design and interactive feedback loops.
My Ownership: Implemented a complete desktop geometry learning app with seven staged modules (2D, 3D, angle, circle, composite shape, sector, stage challenge).
Tech Stack: Java, Swing, object-oriented architecture, modular UI panels, and Javadoc-based maintainability setup.
Measurable Output: Built 7 runnable learning modules, unified score feedback flow, and persistent progress display for productized learning interaction.
Problem: Brain-computer interaction applications need robust signal-to-command conversion under noisy EEG conditions.
My Ownership: Implemented EEG-oriented signal processing and BCI interaction workflow for practical experiment scenarios.
Tech Stack: EEG signal processing, feature engineering, and interaction pipeline implementation.
Measurable Output: Delivered a runnable prototype pipeline covering preprocessing, feature extraction, and interaction output.