Siddhesh Sali

Remote

• Trained convolutional networks for image classification, reaching 97% accuracy in eight epochs.
• Applied transfer learning to object detection, ending with a detection loss below 0.25 after two hours of training.
• Worked through a DevOps curriculum — Docker, Git, cloud compute — and the TensorFlow, Keras, PyTorch, Jupyter and Pandas toolchain.

Mumbai, India

• Built backend and frontend for the festival app and website in Django and Android Studio, for a site taking roughly five million hits a year.
• Worked in a team of ten on Asia's largest science and technology festival, with a footfall over 155,000.

Mumbai, India

• Led 20+ coordinators and workers building stalls, stages and sets across the Student Activity Centre.
• Ran events at Asia's largest college cultural festival — 230+ events, footfall over 143,000, students from more than 1,700 colleges.

Mumbai, India

• Thesis I — Language Guidance for Deep Metric Learning. Multi-stage training that fine-tunes intermediate layers against pretrained language models to produce semantically richer embedding spaces. Guide: Prof. Biplab Banerjee.
• Thesis II — Out-of-Distribution Detection with Vision-Language Features. Extended Maximum Concept Matching from a zero-shot to a few-shot method with prompt tuning. Guide: Prof. Biplab Banerjee.
• Cost-Effective Spin-Coating System, under Prof. Sandip Mondal — design, control and evaluation of a benchtop spin coater.