Qwen Family Preliminary Benchmark Study
An applied research project exploring model size, quantisation, context settings, KV-cache behavior, and inference performance for small local models in scheduled enterprise automation.
Local inference / RAG / computer vision / automation
AI Systems, Automation & Applied Projects
A project archive covering local LLM inference, RAG systems, computer vision, model benchmarking, QR workflows, and automation experiments. The emphasis is on working systems, implementation notes, and the tradeoffs behind them.
Featured Projects
Selected work across classical computer vision, public RAG assistants, local model evaluation, and secure AI infrastructure.
ComfyUI Archive Image Animation
Having fun with LTX 2.3 in ComfyUI to animate archive images of Singapore from the early 1900s to the 1970s.
ONNX + YOLO Face Analytics
Weekend experimentation with ONNX and YOLO computer vision: face boxes, lightweight analytics, and a gender detector that is still a little screwy and needs more work.
PopHistory Singapore History Chatbot
A public knowledge assistant for Singapore history. The retrieval pipeline combines BM25 lexical search, vector similarity search, and a reranker to produce grounded answers through a simple public chat interface.
2012 QR Scanning with ZXing Libraries
A QR decoding experiment using ZXing to locate finder patterns, sample the code grid, handle error correction, and turn camera input into machine-readable payloads.
2012 HAAR Cascade Feature Detection
A classical face-detection experiment using HAAR-like rectangular features, sliding-window scanning, scale pyramids, and cascade classifiers.
How the projects connect
The through-line is practical system behavior: how retrieval quality changes answers, how model configuration changes inference reliability, how local hardware constrains deployment, and how older computer vision workflows connect to current automation patterns.
The pages keep the emphasis on what was built, what was tested, and what remains uncertain.
Research Blog Seeds
- Why Small Local Models Matter for Enterprise Automation
- Qwen Benchmarking: Lessons from a Full Factorial Experiment
- Designing a Prompt Router for Enterprise AI
- RAG Systems: From Embeddings to Retrieval-Augmented Answers
- Local AI Infrastructure: Balancing Cost, VRAM, Context, and Throughput
- From Classical Computer Vision to Modern Multimodal AI
Project Notes
Additional context on the infrastructure, experiments, and research questions behind the project pages.