Akshath Mangudi

I pushed out the first public MVP of Refrakt, marking the transition from a private, evolving system into a shared artifact.
Walkthrough: Refrakt
Public Artifact: refrakt.akshath.tech

I have also started to delve deeper into interpretability-focused work under Iksha, an early-stage research effort.

I won the hackathon hosted by Puch.ai, where I later landed an internship at the company as an AI Engineer Intern, gaining hands-on experience building native AI features, and shipping them to scalable production environments.

Alongside this, I began contribution to open-source projects such as Lighteval, an evaluation framework for LLMs, and sktime, a time series analysis library in Python.

I worked as an Intern at Tectoro, where I learned about the importance of data collection, cleaning, and preprocessing for downstream computer vision tasks.

Towards the end of the internship, I rebranded Re-Implementation into Refrakt, a private repo that was initially scattered into Jupyter notebooks and now a unified platform for reproducible research.

My team and I won the hackathon hosted by Onlinesales.ai, where we built a real-time footfall tracking and analytics platform for retail owners.

I also published the preprint on ViKANFormer, and was shortly acknowledged by another paper that was cited in the ACM Computing Surveys 2025, directly crediting us for the reproducibility we showed in the preprint.

I collaborated with a team to build rgb-to-hyper, a two-stage computer vision pipeline that aimed to reconstruct hyperspectral images from RGB input images for detecting microplastics in water samples.

The project was presented at the university's engineering expo and was later deferred indefinitely due to time constraints and lack of resources.

I completed my internship at Swimlane, where I learned how model decisions interact with trade-offs in latency, accuracy and cost. In parallel, I began a focused research work on ViKANFormer, a benchmark study of KAN-based Vision Transformers.

Several side projects from this period were intentionally abandoned as their scope didn't justify long-term maintenance, pushing me towards fewer but deeper efforts.

I continued re-implementing research papers and took on the role of president of GPUG, which pushed me into collaborating with NVIDIA's Deep Learning Institute to run deep learning and GPU programming workshops to students all over India.

During this period, I secured a Machine Learning internship at Swimlane, gaining hands-on experience with prompt engineering, downstream model deployment and quantization.

Swimlane was my first exposure to ML that was outside my private GitHub repos or Jupyter notebooks.

I began focusing deeply on machine learning through computer vision, where I started to learn about architecture choices and training strategies that were paramount to the success of the model.

Alongside object detection pipelines and delving into LLMs, I started re-implementing research papers directly in notebooks, an effort that later evolved into the earliest versions of Refrakt.

I permanently switched to Linux and moved away from TensorFlow to PyTorch. The outcome of my learning experience was: Summarize4Me, an end-to-end pipeline for extractive summarization of text.

I also built: LLMBot, featuring LangChain and LLMs to parse and answer questions related to research papers attached to prompts.

I began hands-on work with machine learning through Kaggle, focusing on end-to-end model development in Tensorflow. I implemented early predictive models for COVID case trends and stock market predictions, gaining first exposure to data preprocessing, EDA, and model evaluation.

My profile: akshathmangudi

I moved from Seattle to Hyderabad right before my 11th grade and started to prepare for JEE. The period was intense, but I managed to get into VIT-AP as a B.Tech Computer Science student.

0/10 recommended.