Kelvinintech

Problem

Image classification across 133 dog breeds is hard to do from scratch — you need strong features, enough data, and a path to deployment without drowning in boilerplate.

Solution

Built a convolutional neural network with Keras and TensorFlow using transfer learning, plus a Streamlit UI for real-time predictions from uploaded images.

Impact

End-to-end pipeline from training to local deployment, documented on Hashnode (Part 2 of the series; Part 3 benchmarks PyTorch implementations).

Tech stack

Problem

Framework tutorials rarely build intuition for autograd, loss landscapes, and why architectures behave the way they do.

Solution

A hands-on lab of 40+ experiments: custom layers, gradient flow, diagnostics, and visualization — built beside long-form writing.

Impact

Pairs directly with 32+ Hashnode articles in the “PyTorch Zero to One” series so readers can reproduce and extend each idea.

Tech stack

Problem

Genomic and protein sequences are large and awkward to parse — scripts turn into one-off glue instead of reusable pipelines.

Solution

Python utilities with Biopython and NumPy for streaming I/O, batch processing, and modular analysis steps.

Impact

Reusable foundation aligned with software engineering practice and future work in latent-based directed evolution / protein design.

Tech stack

Problem

Students need a way to track terms, courses, and assessments anywhere — even offline — with reminders so key dates don’t slip.

Solution

Built an offline-first Xamarin companion app with a local SQLite database to manage academic terms, course details, and assessment schedules, plus notifications for important deadlines.

Impact

A practical companion to the cloud platform: fast, reliable offline access, consistent data modeling, and a simple UX for staying organized.

Tech stack