Selected projects

Aug – Sep 2026 · AltoTech internship

On-premises LLM Assistant

LoRA · Hybrid RAG · Agent evaluation

Fine-tuned a 30B language model for a 13-tool assistant for building energy management, using a 792-example English/Thai corpus curated from real agent traces.

My contribution

Implemented vector + keyword retrieval over 41 equipment manuals and chunked cross-entropy loss for full 32k-token training inputs on one 80 GB GPU.

Evaluation

Retrieval improved from 8/12 to 10/12 on the reported check. The broader agent suite covered 289 cases for tool selection, arguments, retrieval, and safety. The 12-item retrieval check is a small evaluation sample.

Jun 2025 – May 2026 · Meticuly internship

Production CT Classification Pipeline

DICOM → 2D views → BiomedCLIP + EVA-02 → Quality ranking

A pipeline for the production CraNeXt system that picks the best post-craniectomy skull scan from all the CT series in a case.

Problem

Each case has many CT series: different planes, slice thicknesses, and bone or soft-tissue windows. The first version ranked them only with rules on DICOM metadata and folder names (gantry tilt, missing slices, slice thickness, bone window, pixel spacing, plane). Metadata cannot show what is in the image, so the top pick could be a skull scan without the craniectomy.

What I built

Converted DICOM to NIfTI and made axial, coronal and sagittal 2D views. BiomedCLIP labels the body region (skull/brain, skull and neck, neck, chest, abdomen). For skull scans, a fine-tuned EVA-02 classifies pre- or post-craniectomy, and the ranking now prefers post-craniectomy scans. The output is a ranked JSON file.

Model choice

Benchmarked 8 models (5M–1,012M parameters) on accuracy, speed and size, and chose the smallest, EVA-02 Tiny.

Results

99.4% body-region accuracy (321 of 323 predictions) and 96.1% craniectomy accuracy (99 of 103).

• DICOM
• NIfTI
• BiomedCLIP
• EVA-02