Six peaks, in the order I actually climbed them. A flag goes in at the foot of the
screen as each one comes into view.
2,490 m · peak 01
Health Monitor
Where it started: sensor integration on the ESP32 — MAX30102 heart-rate and MLX90614 infrared
temperature over I²C, read and conditioned straight from the datasheets. Roughly estimated blood
pressure from the pulse waveform, with a small web interface in Blynk to display it all. Some STM32 work
alongside, and enough of the bus protocols to debug them at 2 a.m. with my teammates. Between a FIFO that overflowed, an
MLX90614 that refused to init until I dropped I²C to 100 kHz, and newlib-nano silently
eating my float prints My own blood pressure, meanwhile, hit a solid 250/10.
View on GitHub →
2,790 m · peak 02
Neural Network in C
A neural network in pure C : matrices, forward pass, backpropagation and the training
loop written by hand, no libraries underneath. The Network itself is trained on PIMA Indians Diabetes dataset, and the model is then used to predict diabetes.
I am still traumatied form those segmentation overflow and that one pointer. Yes, I still remember you Mr.Dangling Pointer
View on GitHub →
Bronze · regionals
3,090 m · peak 03
Aarogya
Verifiable professional credentials with the identity left out.
Aarogya proves a doctor holds a valid NMC surgical licence without revealing their name or
registration number. BBS signatures over BLS12-381 run zero-knowledge proofs over the
undisclosed fields, so every presentation is unlinkable — is it real? answered without
leaking who. And is it still valid? lives on-chain: a permissioned Solidity registry of
issuer keys and credential status, where only a hashed ID ever touches the chain. I learned
ZK proofs the honest way: by starting with roughly zero knowledge of them.
View on GitHub →
So, I didn't learn from my stupidity of building NN from scratch in C. So, I built a whole automatic differentiation engine
in C++ from scratch, with Eigen for the operations and autograd for backpropagation. Then I built a GNN and a Diffusion model on top of it.
The best part: the same engine runs both a fraud-detecting GNN and a shape-hallucinating diffusion model. The engine is the product.
The worst part: my tensors captured each other by shared pointer and became so codependent they refused to die. Had to teach an autograd engine
about letting go.
View on GitHub →
🏆 Hackathon winner
A two-mode financial fraud detection pipeline — batch scoring and live transaction screening —
built and shipped under hackathon clock pressure. Most of the model work happened here: CatBoost
carrying the classification, with isotonic regression calibrating the output so the scores mean
what they claim to. It won. The EDA also revealed that, according to the data, all financial
crime in the universe happens on Tuesdays — turns out synthetic data has a weekly routine. Had to
fire day_of_week before it convinced the model that fraud is just a Tuesday problem.
View on GitHub →
The unclimbed peak — and somehow the one I'm most excited to suffer through. Waiting in the fog:
storage engines, the parts of an OS nobody thanks you for, a language of my own, and a baby of two of my projects; tiny device mind elsewhere.
The summit itself stays a secret till the end of this climb: no flag
up there yet. Which is the whole point: a peak you've already reached stops pulling you upward. This
is the one that gets me out of bed (the coffee helps).
Route in progress ↗