Deep learning research engineer rebuilding frontier AI architectures from scratch in PyTorch — LLMs, latent diffusion, multimodal, video.
- 11 from-scratch projects spanning LLMs (LLaMA-3, DeepSeek-V3, FusionLLM, GPT-2, TranslationLM), generative vision (Stable Diffusion, DCGAN, VAE, CycleGAN), multimodal (PaliGemma-style VLM), and video understanding (ST-GCN action recognition).
- Single-GPU feasibility: engineered to run on A100, RTX 5090, RTX 6000 Ada, RTX 3090, P100, T4 with BF16, Flash Attention 2, gradient checkpointing, and torch.compile.
- Faithful paper reproductions: MLA with absorption trick, aux-loss-free MoE, Multi-Token Prediction, Min-SNR, KL annealing, AdaIN conditioning.
| Project | Status | Link |
|---|---|---|
| Stable Diffusion 1.x (860M UNet) | Trained 42 epochs on 2x RTX 5090 | repo |
| VisionLanguageModel (PaliGemma-style) | Trained end-to-end on COCO captions | repo |
| FaceAgingCycleGAN (AdaIN) | 31 epochs on IMDB-WIKI | repo |
| FaceGenerationVAE (beta-VAE) | 50 epochs on CelebA | repo |
| DCGAN-Face-Generation | 50 epochs on CelebA | repo |
| GPT-From-Scratch | Trained on Tiny Shakespeare | repo |
| TranslationLM (EN->IT) | 20 epochs on OPUS Books | repo |
| Project | Status | Link |
|---|---|---|
| DeepSeek-V3-Lite (MLA + MoE + MTP) | Architecture and pipeline ready; pre-training not started | repo |
| LLaMA-3-Lite (515M) | Architecture and pipeline ready; pre-training not started | repo |
| FusionLLM (MLA + GDN + MoE + MTP) | Framework and tests ready; pre-training not started | repo |
| ActionRecognition (ST-GCN) | Pipeline and tests ready; NTU benchmark not started | repo |
- Portfolio site: atandra2000.github.io/mycv
- LinkedIn: linkedin.com/in/atandrabharati
- Weights and Biases: wandb.ai/atandrabharati-self
- Kaggle: kaggle.com/atandrabharati
- Comet: comet.com/atandrabharati