Krushna Parmar
Just a curious guy who likes building things that actually work.
I'm Krushna. I like to know how things work, all the way down. How a model learns. How memory moves on a GPU. How some machine from a hundred years ago actually ran. So I build small, sharp tools, take things apart, and lose track of time doing it.
I'm doing my M.Tech in Machine Learning right now, and most of what I make is open source. I build fast, ship under pressure, and care more about making the hard thing clear than talking about it.
Settling into my M.Tech, tinkering with post-training pipelines, and rebuilding an old machine on the side.
Selected Work
Eco-Sattva
MERN-stack carbon-footprint tracker with ML integration for sustainability metrics.
AR Campus Navigation
ARCore-based indoor campus navigation without GPS, using geolocation and QR recognition.
Chat App
IoT LocalHost chat over ESP32 and Bluetooth BLE with a custom UI.
Feel free to explore my GitHub for more projects. Most of them are open-source.
Experience
Parsewave
Frontier AI Engineer
Jun, 2026 - Jul, 2026
Remote
- Built tasks and datasets for post-training optimization of large language models
- Authored Harbor-format tasks and Terminal-Bench evaluations to probe how models behave in real shells
- Reviewed and benchmarked model runs, then folded the findings back into sharper tasks
- Built pipelines that generate tasks for post-RL optimization at scale
CHARUSAT
Machine Learning Research Intern
Dec, 2025 - Apr, 2026
Anand, India
- Wrote CUDA kernels to accelerate graph algorithms on large, messy real-world graphs
- Designed and built four original CUDA algorithms (GPU-BiDS, GPU-A*, GPU-BiD-A*, and GPU-BatchPPSP) for bidirectional and heuristic shortest-path search
- Co-authored MARGA: GPU-Accelerated Bidirectional and Heuristic Shortest-Path Search on Real-World Graphs Using CUDA (under review)
- Built the code-generation backend for HuskML, so you design a neural net visually and export working Keras or PyTorch code
HuskML
Founder
Jul, 2025 - Oct, 2025
Remote
- Founded HuskML, a visual, browser-based neural-network designer that exports production-ready Keras / TensorFlow code
- Built open-source educational resources with cross-platform compatibility
- An ambitious build I couldn't fully execute solo. A failure worth sharing, and worth continuing
BISAG-N (MeitY - Govt. of India)
ML Research Engineer Intern
Jun, 2025 - Sep, 2025
Gandhinagar, India
- Built the data pipeline for multi-temporal Sentinel-2 imagery over the Bihar Darbhanga and Surat floodplains, engineering NDVI, slope, curvature and land-use features
- Evaluated Random Forest, XGBoost, SVM, Naive Bayes and Logistic Regression for flood-risk classification; deployed XGBoost to production at a realistic 71% accuracy on noisy field data
- Built three QGIS plugins in PyQt5 to automate flood-boundary extraction from DEM rasters, point-level inundation likelihood, and real-time flood-extent rendering from ML outputs
DailyHijacks
R&D Engineer (Freelance)
Apr, 2020 - Jan, 2023
Remote
- Implemented Telethon API and built Python Telegram bots deployed on Heroku
- Ran web development, SEO optimization, graphics, and content across the company's site and social channels
- Researched and wrote SEO-optimized blogs on digital-marketing and technology trends
Education
M.Tech, Machine Learning
Dhirubhai Ambani University
B.Tech, Artificial Intelligence
CHARUSAT
Fun Facts
- Two national hackathon wins so far.
- A Governor of Gujarat once handed me a quiz award.
- I placed 24th at a state chess tournament.
- I rebuild ancient tech just to see how it ticked.
- Badminton is how I reset.
- I get way too excited about how memory actually works.
- Most of my code lives out in the open, on GitHub.
Thoughts & Insights
A running notebook. What I learn while building, in computer vision, ML, and the odd rabbit hole. Notes to my future self that happen to be public.
Read my thoughtsCurated
A few things worth your time. The stuff I keep coming back to.
- read The Illustrated Transformer Still the clearest mental model for attention.
- repo tinygrad A whole deep-learning framework you can actually read.
- read Spinning Up in Deep RL Where post-training stopped feeling like magic to me.
- docs CUDA C++ Programming Guide The book I keep open while writing kernels.
On Repeat
