Design pattern extraction and imply into statistical model

The core idea is that CAD assemblies encode implicit design knowledge — how parts relate, what tolerances recur, which sub-assemblies appear together. A Bayesian model treats these as latent variables and learns a probabilistic grammar of design patterns.

Design pattern extraction animation

Automation based on CAD kernel 3D model generation & 2D drawing generation

Mostly CAD software (Solidworks, NX, Creo, Catia, Fusion 360, Onshape etc) use the same couple of underlying CAD kernels. Parametric CAD modeling relies on sketch and extrude operations to define 3D shapes. These sketches are typically defined with constraints and dimensions that support quick edits and design variations. The Framework here bridges with the CAD software kernels and does customization across all the model parts automation utilizing the full potential of hardware.

CAD kernel automation animation

Engineer feedback driven chatbot, intelligence rooted in domain terminology.

Tailored specifically for the Engineering industry. Enhance the team's productivity and simplifies the query resolution pipeline. Simplify based on complex questions breaking down to steps with actionable fixes. Unified workspace enabling peer collaboration integrated with an adaptive bot that evolves from the history of tangles solved before. Mechanism of the framework shaped based on teams approach.

Engineer chatbot screenshot

Digital twin drone (vision-controlled autonomous navigation) inside NVIDIA omniverse for control of systems.

Digital Twin of an industrial-grade drone, trained via Proximal Policy Optimization reinforcement learning and deployed on NVIDIA Omniverse, enabling the dynamicality on control of systems (industrial monitoring, factory planning, simulation).

Engineer chatbot animation

Physics engine (2D based) integration with CAD software

2D physics engine with simple interface (game like feel) to interact with unfolded version of the existing CAD parts to customize and linking them with CAD systems to automate the real CAD part, subassemblies, main assembly (sequential cad modeling process). Simply talking to the CAD software API’s with JSON.

Physics engine integration screenshot

Parking-Spot-Detection Signal-Image-Video

Applied data augmentation (rotation, brightness shifts) to improve CNN generalization across varying lighting and weather conditions. Integrated OpenCV for frame pre-processing and ROI extraction, enabling optimized real-time inference on live camera feeds. Project supervised by Professor Francesco De Natale.

Feature Detection (Computer Vision)

Benchmarked Lucas-Kanade, SIFT, and ORB across sequences, analyzing trade-offs in keypoint repeatability, descriptor discriminability, and computational cost. Added homography-based geometric verification to filter outlier correspondences, improving robustness for pose and motion recovery. Project supervised by Professor Nicola Conci.

Deep Learning with CLIP on RefCOCOg

Fine-tuned CLIP encoders using contrastive loss to align region-level visual embeddings with referring expression descriptions. Evaluated the joint embedding space via recall@K metrics and t-SNE visualization to assess cross-modal alignment quality. Project supervised by Professor Elisa Ricci.

Language Model Optimization

LLM and regularization techniques with RNN and LSTM model with optimizers for improving the perplexity following with fine-tuning of BERT model for improving sequence labeling and text classification task. Project supervised by Professor Giuseppe Riccardi.