Thura Win Kyaw

Internship Project · AIdanBio In Progress

AIdanMED

Label-Efficient VLM Post-Training for Faithful, Grounded Medical Reasoning

Medical VLMGRPOSFTGrounded ReasoningBounding Boxes

Built during my internship at AIdanBio, this project trains a VLM with GRPO to produce reasoning together with a bounding box that must remain faithful to the model’s own reasoning, targeting label-efficient and grounded medical reasoning.

Status: In progress. SFT cold-start complete; GRPO post-training is running on 8×A100 GPUs. No evaluation results reported yet.

Research Setup

Setting: Medical vision-language
Objective: Faithful grounded reasoning
Output: Reasoning + bounding box
Post-training: SFT cold-start + GRPO
Compute: 8× NVIDIA A100
Current stage: GRPO training

Built at AIdanBio In Production

ChatPDB

Protein Structure Visualization Assistant

LLMRAGPyMOLFastAPIVector DBPython

An intelligent assistant that interprets natural language queries to visualize 3D protein structures from the RCSB PDB. Integrates LLMs with a RAG pipeline for semantic understanding of protein documentation, and PyMOL for interactive rendering — enabling biomedical researchers to query structural data conversationally without writing visualization scripts.

System Architecture

User Query natural language

LLM Parsing intent & entity extraction

Vector Retrieval protein metadata & docs

PyMOL Generation command synthesis

3D Visualization interactive output

Independent Research In Progress

PhysHolo

Physics-Informed Holographic Phase Reconstruction

Computational ImagingDifferentiable OpticsAngular SpectrumPyTorchU-Net

Explores whether a differentiable optical forward model can improve lightweight neural phase reconstruction from synthetic holograms. The study focuses on comparing classical back-propagation, a compact U-Net, and the same U-Net trained with a physics-consistency loss, with measurement noise and unseen propagation distances as robustness evaluation targets.

Status: In progress. Current focus: differentiable Angular Spectrum propagation, simulator validation, and reconstruction baselines.

Research Setup

Initial scope: Synthetic 2D phase reconstruction
Forward model: Angular Spectrum Method (ASM)
Learning: U-Net + physics-consistency loss
Evaluation: MAE, RMSE, PSNR, SSIM + hologram consistency
Robustness: Noise + unseen propagation distances

Independent Research Preliminary results

MoonDPO

Memory-Efficient DPO for Hallucination Mitigation in Small VLMs

Moondream2 1.9BDPOLoRACOCOPOPECHAIR

Applies LoRA-based DPO to Moondream2 (1.9B) using preference pairs automatically generated from COCO annotations, without human or API labels. The DPO objective is derived and implemented from scratch, and the full pipeline is reproducible on a single consumer GPU (16–24 GB).

**Preliminary results · Single seed** 5,000 images → 1,945 pairs · LoRA r=32 · 3 epochs POPE: 500 questions/split · CHAIR: 200 images
Metric	Baseline	MoonDPO
POPE F1 · Random ↑	0.919	0.921
POPE F1 · Popular ↑	0.900	0.915
POPE F1 · Adversarial ↑	0.866	0.871
POPE Recall ↑	0.932	0.928
CHAIR_i ↓	0.132	0.031
CHAIR_s ↓	0.215	0.035
Objects / caption	1.86	1.12
↑ Higher is better · ↓ Lower is better
Fewer objects per caption indicate a caption-richness trade-off. Training/evaluation image-overlap checks, multi-seed evaluation, and VQA regression remain pending.

Research Setup

Base model: Moondream2 (1.9B)
Adaptation: LoRA + DPO
Preference data: COCO-derived pairs
Compute: Single consumer GPU
Evaluation: POPE + CHAIR

Undergraduate Researcher in Multimodal AI

AIdanMED

ChatPDB

PhysHolo

MoonDPO

AI/ML/DL Engineer Intern

AI Software Engineer Intern

B.E. in Electrical Engineering