Ekalabya Ghosh

Open-Weight Real-Time Decision Model for Text, Images and Video

Viveki-0.1-20B is a general-purpose decision model built as an adapter on a ~27B-parameter base model. Instead of generating text, it reads out a score for each option of a question you define, so a single forward pass returns a probability per option and no tokens are generated. Its forward pass exits early at an internal readout layer, so only ~20B parameters execute per request. It accepts text, images and video frames together and can answer several questions about the same input at once. For video, it streams decisions at roughly 1 Hz using a persistent KV cache over a sliding window, which makes it suited to event and state detection (for example, a goal being scored or a fight breaking out) with a caller-calibrated alert threshold. It is released on Hugging Face under Apache 2.0.

The same readout that decides yes or no also shows where in the frame the answer comes from. Described only in plain text ("Is a forklift carrying a tall stack of three or more crates?"), the state lit up the crate stack on real 1920x1080 factory CCTV, with no boxes, no generated tokens and no training for the task. The map is a coarse heatmap (32 px cells) that is only trusted while the model answers yes, and the demo keeps its one miss, when the forklift cab was hidden. Scene-level states worked; fine details like an open guard gave weak maps. It runs at about 4.8 s per frame, so it produces maps rather than driving the real-time loop.

❮❯

Novel Deep-Learning Pipeline for Explainable Diagnosis of Thoracic Diseases

Current CNN-based models for diagnosing thoracic diseases lack clinically aligned interpretability and perform suboptimally (AUC ~0.7) on certain pathologies. The kernel size of the CNN limits its ability to captured global patterns. RadFusion introduces a multimodal transformer framework trained on ~600k image-report pairs (combining the MIMIC-CXR and ChestX-Ray14 datasets) to address these problems. The project introduces a novel augmentation technique called CHEF (Conditional Histogram Equalisation and Feature) augmentation that dynamically blends different augmentation techniques during training to enhance feature generalization and reduce class bias. The project also introduces a novel disease specific attention mechanism that gives a query vector to each disease to capture effective correlations between features and specific diseases. The model is especially useful in busy hospitals where radiologists can supervise multiple models simultaneously to increase their productivity.

Self-Hostable Backend for Easy Deployment and Monitoring of Models on AWS

Eezy-ML is a self-hostable backend solution designed to simplify the deployment and monitoring of machine learning models on AWS. It comes with a custom python SDK for setting up projects and seamlessly integrates with GitHub for easy version control. It has support for custom UIs that can be configured from the project source code for A/B testing, benchmarking and other use cases. The server has a very low latency of ~150ms for deployed SLMs. The servers have built-in support for horizontal scaling and load balancing. The project is built with a focus on user-friendly interfaces and robust performance, making it an ideal choice for developers looking to deploy machine learning models without the hassle of managing complex infrastructure.

Autonomous Desktop Agent for Managing Memories

Ghost is an AI-powered personal memory assistant, built for use cases where AI needs a lot of context about the user's workflow, preferences, or needs. It looks at the screen of the user, process the screen recordings in small chunks and builds a vector database (FAISS) with the information it sees to build rich context about the user. The videos are pre-processed to reduce the length and resolution of videos while keeping the key information visible. Afterwards, the videos are fed into the gemini 2.5 flash lite model that generates memory fragments. These memory fragments can form connections with documents, folders, github-commits and other memories. To make it feel like an actual assistant, Ghost can be connected to apps like GitHub. The project is entirely open-sourced. It also has additional applications in ADHD care.