Computer Vision · Deep Learning · Agentic AI · Edge ML
M.Eng. in ICT — Asian Institute of Technology, Thailand
I work at the intersection of computer vision and physical systems—where a model's output has real-world consequence. My thesis detected and measured individual oil palm trees from drone imagery across 8 altitude levels using YOLOv8, YOLOv11, SAM 2.1, and Mask R-CNN. I built an agent-in-the-loop annotation pipeline that reduced labeling time by 80% and a generative tiling algorithm that simulates multi-altitude flights from a single drone pass.
Recent focus: agentic AI systems that are auditable and constrained, edge vision models under strict latency and power budgets, efficient training of LLMs and small models, and local-first tools for knowledge work.
Open to AI Engineer, Computer Vision, Remote Sensing ML, and Agentic AI roles in Thailand, India or remote.
- Deep Reinforcement Learning (HuggingFace RL Course)
- Langchain Academy Course
- Kaggle Competetions / Hackathons
I built Dopamine.Diet: a local-first productivity system that hijacks your brain's love for checkboxes, streaks, and green squares — and redirects it toward deep work.
- Live Blog
- Website: Dopamine.Diet
Multi-agent orchestration system with 9 specialized agents, 6 immutable writing laws, and 90% confidence gate for hallucination control. Built during thesis at AIT to automate research documentation and inference.
- Repo: Thesis-Agent-Kit
Reproduced OpenAI's Parameter Golf challenge: H100 training on RunPod with torchrun, FineWeb dataset, and continuous val_bpb tracking. Focus on reproducibility under strict wall-clock constraints.
- Demonstrates large-scale training orchestration and compute efficiency
- Repo: OpenAI-Parameter-Golf
Image-to-text retrieval on XR2 Gen 2 proxy (edge device). Achieved Recall@10 ≈ 0.73.
- Full pipeline: model selection, ONNX export, hardware profiling, dataset curation, inference optimization
- Demonstrates end-to-end edge ML deployment
Zero-framework Jekyll blog with dark mode, image lightbox, giscus comments, RSS, keyboard navigation, and reading progress indicator. Built via AI pair programming; MIT licensed, open source.
One-click Obsidian-to-Jekyll converter with automated frontmatter, image copying, and intelligent filename generation. ~8KB, zero dependencies.
- Bridges knowledge management (Obsidian) and blogging (Jekyll) workflows
- Repo: obsidian-jekyll-publisher
Browser console utility for extracting structured data from HTML tables and forms. Zero external dependencies; designed for rapid prototyping and data collection.
Detection, counting, and canopy biometry of individual oil palms from UAV imagery at multiple ground sample distances (0.03–0.20m).
- Multi-model comparison: YOLOv8, YOLOv11, Mask R-CNN, SAM hybrids
- Metrics: precision, recall, F1, IoU, crown geometry errors across 8 altitude levels
- Agent-in-the-loop annotation (80% labeling time reduction) + generative tiling for synthetic multi-altitude data
A technical series documenting the full research and deployment journey:
- Production-grade AI systems: Converting research prototypes into testable, deployable, and maintainable systems with clear evaluation metrics
- Edge & on-device vision: Geospatial ML and vision models that run under strict latency, power, and memory constraints
- Building Projects: Building productivity systems utilising OCR/TTS/RAG-Chatbots that keep data local and enable fast context-switching between projects
- Reliable, auditable agents: Agentic AI with constraint-based reasoning, explicit control flow, and verifiable outputs (not just creative generation)
"The tension you feel is not a problem to solve. It is the process."