Deep RL for IC Macro Placement
B.Tech Project · IIT Roorkee, under Prof. G. N. Pillai
Macro placement framed as sequential decision-making. A dual-branch convolutional encoder
(ResNet-18 global context + local fusion) reads pixel-mask state — position feasibility,
wire cost, occupancy — and feeds policy and value heads, replacing purely hypergraph-based
approaches with spatial visual reasoning over the canvas.
On Ariane RISC-V, DDPG reached 155,811 HPWL against PPO's 179,802; on the
heterogeneous Adaptec-1 SoC, PPO led with 914,196 HPWL. Both converged stably with
effective overlap avoidance and congestion control.
PyTorchPPODDPGGymhMETIS
View repository →
Chandrayaan Moon Mapping Silver
Inter IIT Tech Meet 11.0 · ISRO problem statement
Applied ESRGANs to upscale lunar imagery from 128×128 to 2048×2048, then implemented
image stitching to assemble the tiles into coherent high-resolution lunar maps.
Won a Silver Medal against the other IITs.
PyTorchESRGANOpenCV
View repository →
S-Wave — Learned Sparse Wavelets
Signal representation learning · grew out of my wavelet research at IIT Roorkee
Learning sparse wavelet representations tailored to a specific signal, instead of picking a
fixed basis off the shelf. An autoencoder-style setup trains a FilterConv model
to predict low-pass filters while minimising reconstruction loss through a differentiable
DWT, so the basis itself is learned end-to-end.
Builds on Adaptive Wavelet Distillation from Neural Networks through Interpretations
(Ha et al., 2021), currently implemented for 1D signals.
PyTorchWaveletsDWTSignal Processing
View repository →
Weld Defect Classification
Analytics Challenge · IIT Bombay Techfest — Top 5 finalist
Predictive defect classification over 800,000+ welder data points using Random Forest,
XGBoost and SVM, reaching a weighted F1 of 0.90. Included EDA and t-SNE feature
visualisation to understand class separability.
scikit-learnXGBoostt-SNE
View repository →
Stock Forecasting with Time2Vec
Time-series deep learning
Transformer and LSTM models trained on 40 years of market data using Time2Vec embeddings.
Achieved R² up to 0.982 (Apple), 0.978 (IBM), 0.970 (Goldman Sachs) and 0.963 (Coca-Cola).
TensorFlowKerasTransformers
View repository →
RAG over Project Gutenberg
Retrieval-augmented question answering
A question-answering interface that answers from story plots, built as a retrieval-augmented
generation pipeline over the Gutenberg corpus.
PythonRAGLLMs
View repository →
Audio Classification
Freesound Kaggle Challenge · UrbanSound
Multiclass–multilabel classification directly from .wav audio for the Freesound
challenge, plus an MFCC-extraction and LSTM pipeline on the UrbanSound dataset.
MFCCLSTMPyTorch
View repository →
