Data Scientist
I rebuild published models from the paper — the transformer, BERT, ViT — tensor by tensor in PyTorch, all three on Bengali data, all three trained end to end on one M1 laptop with 16 GB of RAM. Reading a paper and reimplementing it turn out to be very different kinds of understanding. By day I build forecasting and agentic systems at Sigmoid.
Re-engineered daily demand forecasting across 9,200 machine × product series — full run from ~3 hours to ~3 minutes. Plus an MCP proof of concept driven from three model backends.
Built the feature-engineering pipeline behind a four-agent RAG recommendation engine, and worked on one of those agents — live since June 2025, ~80% accuracy in its core categories. Earlier, price elasticity and promo baselines for a global CPG manufacturer.
Measured which levers — price, promo, media, retail execution — actually move sell-through for a global spirits manufacturer, so spend could be aimed where it pays.
papers-from-scratchSOTA architectures rebuilt in PyTorch
Three papers replicated tensor by tensor on Bengali data — no nn.Transformer, no nn.MultiheadAttention, no pre-built HuggingFace models.
BERT: Pre-training of Deep Bidirectional Transformers for Language UnderstandingDevlin et al., 2019
A 7.5M-parameter Bengali BERT, pre-trained from nothing on a laptop, that scores above published mBERT and IndicBERT on Bengali news classification.
Attention Is All You NeedVaswani et al., 2017
The original transformer rebuilt component by component, then trained to translate English into Bengali on one laptop.
An Image Is Worth 16x16 WordsDosovitskiy et al., 2021
A ViT trained on 1,935 photographs of Bengali terracotta temples — a dataset I had to build first — to test the paper's claim about pre-training. It failed, exactly as predicted.
Haldia Institute of Technology
CGPA 9.05. Affiliated to Maulana Abul Kalam Azad University of Technology, West Bengal.
Jhantipahari High School
Jhantipahari High School
A 7.5M-parameter BERT pre-trained from random weights on Bengali Wikipedia in 28 laptop-hours, then fine-tuned to 86.5% on IndicGLUE sna.bn — above published mBERT and IndicBERT.
An ~11M-parameter Transformer built tensor by tensor on a single MacBook — the full math, the build order, and the findings only constrained hardware surfaces.
A ViT trained on 1,935 photographs of Bengali terracotta temples — a dataset I had to build first — to test the paper's claim about pre-training. It failed, exactly as predicted.
Everything here is something I have shipped with or trained with.
Languages
Deep learning
Generative and agentic
Modelling and statistics
Data and cloud
![86.5% on IndicGLUE sna.bn, above mBERT and IndicBERT — from 7.5M parameters trained 28 hours on one 16GB M1. On the right, the six-layer encoder stack: a [CLS] A [SEP] B [SEP] pair goes in at the top, a pooled [CLS] classifier reads it out at the bottom.](/_astro/bert_cover.ChghBB-Z_Z16YI7L.webp)
