Project Intern — Samsung PRISM
Dec 2025 — PresentSamsung R&D Institute India · Bengaluru
- Developing EmoCapNet, a 254.8M-param multimodal vision-language model for 7-class emotion-aware image captioning — fusing ViT-Base + GPT-2 via cross-attention and VAD-FiLM conditioning with multi-task learning on a 67K-caption emotion-augmented Flickr30k corpus.
- Distilled the teacher into a TinyCLIP-ViT + DistilGPT-2 student (112.5M params, 2.3× smaller); achieved METEOR 22.1, CIDEr 0.363, and emotion-controllability correlation r = 0.81.
- Engineered on-device deployment via INT8 dynamic quantization (custom Conv1D→Linear rewrite), compressing ~1GB → 153MB (6.7×); built greedy KV-cache decoding from scratch (524ms/caption CPU) with an ONNX/TorchScript export path.