Skip to content
View AbhishekAQ's full-sized avatar

Highlights

  • Pro

Block or report AbhishekAQ

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
AbhishekAQ/README.md

Abhishek Kumar

I build machine learning models and the Python services that serve them. I have been at Ipseity since 2023, as an intern from February and full time from August.

Most of my work there is DART (Diagnostic Algorithm for Reflex Testing), where I built the CNNs that classify pathological features in medical imaging. I also built a rule-based prescriptive system for antibiotic resistance, which takes a pathogen identification, its resistance profile and the patient's own data and returns an antibiotic recommendation. Both run behind async FastAPI services on PostgreSQL and MongoDB. Auth is OAuth2 with JWTs, and there are WebSockets where the data has to be live.

Projects

probabilistic-pde-discovery recovers partial differential equations from image sequences using SINDy. Every recovered term carries a bootstrap confidence interval and a Bayesian MAP credible interval, and ensemble SINDy scores how often a term survives resampling, which is what rules out the spurious ones. The experimental frames come in unregistered, so I align them with optical flow before anything else runs.

In scientific-image-forgery-detection I treated forgery detection as pixel-level segmentation rather than image-level classification, because the manipulated regions are small. I implemented UNet and DeepLabV3+ and compared them over ResNet50 and ResNet101 encoders and an untrained one, with BCE, Dice and a combination of the two. Scored with object-based F1 and Hungarian matching on the Kaggle RecodAI dataset.

music-bot-pg is a Telegram bot that streams YouTube audio into group voice chats. Two queues, one for what people ask for and one for what it picks itself when nobody is asking. It leaves when the voice chat empties. Runs in Docker.

Background

MSc Machine Learning and Data Mining, Université Jean Monnet, Saint-Étienne, 2024 to 2026, overlapping the Ipseity work. BTech Computer Science Engineering, Guru Gobind Singh Indraprastha University, Delhi, 2019 to 2023. In 2022 I interned at BHEL on unsupervised change detection in satellite imagery, PCA then K-means, written in C++ with OpenCV.

Based in Delhi.

Popular repositories Loading

  1. fatal-crash-analysis fatal-crash-analysis Public

    Data mining study of US fatal traffic crashes, 2015 to 2022. Clustering, correlation analysis and predictive modelling in R.

    R

  2. Leetcode-NeetCode-150 Leetcode-NeetCode-150 Public

    Python solutions to the NeetCode 150, with complexity notes and a spaced-repetition tracker.

    Python

  3. music-bot-pg music-bot-pg Public

    Telegram bot that streams YouTube audio into group voice chats. Dual queue, inline controls, containerised with a health check endpoint.

    Python

  4. probabilistic-pde-discovery probabilistic-pde-discovery Public

    Discovering partial differential equations from image sequences with SINDy, with bootstrap, Bayesian and ensemble uncertainty quantification on every coefficient.

    Python

  5. scientific-image-forgery-detection scientific-image-forgery-detection Public

    Pixel-level forgery detection in scientific images using UNet and DeepLabV3+, evaluated with object-based F1 via Hungarian matching.

    Jupyter Notebook

  6. AbhishekAQ AbhishekAQ Public

    Profile