Computer Science Undergraduate • Transformers • Diffusion Models • Deep Learning
I build models from first principles — implement papers, debug training, and optimize GPU throughput. learn deeply → build real things → share.
- Transformers / Attention (encoder–decoder, masking, multi-head attention, cross-attention)
- Training engineering (experiment tracking, checkpointing, validation pipelines)
- Performance (GPU training on A100, measuring BLEU/perplexity, not vibes)
- Clean implementations (readable code, reproducible training)
Encoder–Decoder Transformer (Vaswani et al., 2017), built end-to-end in PyTorch.
Includes multi-head attention, positional encoding, cross-attention, and masking.
- Trained on the IITB English–Hindi dataset (~1.6M parallel sentences)
- BLEU: 44 · Perplexity: 22.09, in ~45 minutes on an A100
- Goal: correctness + clarity + engineering discipline
Forward diffusion and reverse denoising process for generative image modeling.
- UNet-based architecture for noise prediction across diffusion timesteps
- Iterative sampling pipeline for image generation from Gaussian noise
- Deployed as an inference service with FastAPI, containerized with Docker
Distributed microservice-based platform for automated resume analysis and job recommendations.
- Built with FastAPI, PostgreSQL, Redis, Celery, Qdrant, and MinIO
- Asynchronous processing pipelines using Celery workers for large-scale background tasks
- Vector-based candidate–job matching using AI embeddings and Qdrant for semantic retrieval
- Secure REST APIs with JWT authentication, rate limiting, caching, and Alembic migrations
- Containerized and orchestrated with Docker Compose
- correctness > cleverness
- measurable performance (not vibes)
- simple baselines → controlled experiments
- understand internals (attention, padding, scaling, initialization)
- LinkedIn: https://www.linkedin.com/in/shubham-patel-928132288/
- Email: sp710812gs@gmail.com