I work on software systems and applied machine learning. So far I have built an offline emergency messaging app and a trading research tool, and I have trained models on medical scans and brain signals.
Before I trust a result, I try to break the conditions it depends on.
With ResQNet, I asked what an emergency app can still do when the network is gone.
With Trading Lab, I test whether a trading rule still beats holding once every trade pays a cost.
With my medical imaging research, I check that the model is tested on people it has never seen.
०२
These are five projects I have built or researched. Each one has a full write-up that explains how it works and what is still unfinished.
What can an emergency app still do when the network is gone?
An Android prototype that passes SOS messages between nearby phones when there's no internet, along with the problems that remain.
Read the full case studyIllustrative relay path between connected devices. No measured range or delivery reliability is implied.
Does a trading rule still beat holding once fills wait for the next open and every trade pays a cost?
A Python workbench that scores trading rules against buy and hold, with next-open fills and a cost on every trade.
Read the full case studyConceptual research notebook. This drawing represents the evaluation process, not measured returns.
How much of a high accuracy score comes from the model, and how much from the split?
A medical-imaging benchmark that shows how splitting scans by patient instead of by slice changes the reported accuracy.
Read the full case studyIllustrative patient grouping. All slices from one person stay in a single evaluation partition.
Can a classifier learn the condition rather than the person?
A signal-processing and classification study on EEG recordings, with subject-level evaluation.
Read the full case studyConceptual signal illustration, not a patient recording. The research method filters signals before extracting features and comparing classifiers.
Can a distributed bug be replayed exactly, every time?
An early database project in Go. It starts with Raft and a deterministic simulator, before a networked, replicated key-value store.
Read the full project noteIllustrative three-node target architecture. The current project begins with deterministic simulation.
०३
Solo project · Go
I am building a replicated key-value store in Go. I started with a deterministic Raft simulator, so that any failing run can be replayed exactly.
Independent research · Python
In August I built an engine that tests trading rules on daily prices. In October I wrote a draft protocol for opening-range breakouts on Nifty 50 stocks, with a dataset registry and a holdout set that tuning never touches.
Mobile and systems engineering
I built an Android prototype that relays SOS messages between nearby phones without a server. The later versions move the backend onto a self-hosted server.
NIT Rourkela · EEG
I worked with EEG recordings during my internship. I cleaned the signals, extracted features, and compared four classifiers on data split by subject.
Computer vision · Advisor-guided
I benchmarked 36 pipelines to show how splitting scans by patient instead of by slice changes the reported accuracy. I also audited the dataset for duplicate images.
Frontend
I have built animation-heavy websites for clients using GSAP, Three.js, and Lenis.
०४
These are the languages and tools I use most. They change from project to project.
०५