01Jan 2025 - Present
Machine Learning Engineer
AiLusion, Merosa Technologies Pvt. Ltd
Built and fine-tuned large-scale VLMs and diffusion models with PyTorch, Flux-Dev, and CLIP-style embeddings. Improved training and retrieval workflows with mixed precision, distributed evaluation, and targeted adaptation techniques including LoRA, QLoRA, and DreamBooth.
01What I did
- 01
Model training and adaptation
Trained and fine-tuned large VLMs and diffusion models (including Flux-Dev) in PyTorch, choosing the lightest adaptation method that meets the quality bar: LoRA and QLoRA for parameter-efficient tuning, DreamBooth for subject-specific personalisation.
- 02
Training efficiency
Used mixed-precision training and distributed evaluation to shorten experiment cycles, so more ideas could be tested per week on the same hardware.
- 03
Retrieval and embeddings
Built and tuned CLIP-style embedding workflows for multimodal retrieval, improving how well text and image queries match the right items.
- 04
Evaluation
Set up repeatable evaluation so model changes could be compared on the same footing before they reached a product.
02Impact
- Faster experiment turnaround through mixed precision and distributed evaluation
- Lower fine-tuning cost using parameter-efficient adaptation
- Better retrieval quality from improved embedding workflows
03Tech
PyTorch, Flux-Dev, CLIP, Diffusers, LoRA, QLoRA, DreamBooth