Kelvinintech

Problem

Developer dashboards that call multiple APIs often fail as one unit — one broken source blanks the whole page — and LLM demos rarely show grounded retrieval before generation.

Solution

A .NET 8 Razor Pages dashboard with independent cards (Weather, HN, GitHub, Crypto, Daily Challenge) plus Ask DevPulse: real RAG over recent HN + GitHub chunks (embeddings + in-memory similarity), then a grounded LLM answer with citations.

Impact

Copyable mid-level ASP.NET pattern — typed models, HttpClient services, DI, CardResult<T> isolation, and retrieve-first RAG that still surfaces sources if the LLM is down — aimed at an Azure App Service demo.

Tech stack

Problem

Job seekers juggle applications, interview dates, recruiter contacts, and outcomes across tools that don’t talk to each other — usually messy spreadsheets.

Solution

Offline-first Windows desktop app (WPF + ASP.NET Core) that organizes search rounds, tracks interview stages, estimates offer probability, exports CSV reports, and syncs SQLite data to a JWT-secured cloud API on SQL Server.

Impact

End-to-end product thinking: domain modeling for a real workflow, client/server separation, offline resilience, and testable business logic — not just a CRUD demo.

Tech stack

Problem

Most breed-classifier demos lean on transfer learning — it’s easy to skip the hard parts of designing capacity, regularization, and a real train/val/test loop on a small dataset.

Solution

Built and trained a custom CNN (no transfer learning) on the Kaggle Dog Breed Identification dataset to classify Scottish Deerhound, Maltese, and Bernese Mountain Dog — data prep, splits, L2-regularized architecture, Colab training, and a Streamlit app for upload → breed + confidence.

Impact

Full training-to-deployment loop that builds intuition around convolutional feature extraction, capacity control, and overfitting — not just a notebook glued to a pretrained model.

Tech stack

Problem

Some tarantula species look alike — morphology alone doesn’t tell you which are genetically close, or how far apart look-alikes really are.

Solution

An R bioinformatics pipeline that pulls COI DNA barcodes for closely related tarantula species, aligns sequences, computes pairwise genetic distances, and builds a neighbor-joining phylogenetic tree.

Impact

Uses mitochondrial COI barcode sequences to compare look-alike tarantula species, quantify genetic distance, and visualize relatedness — grounding sequence alignment and distance concepts that scale into latent-space / directed-evolution work later.

Tech stack

Problem

Students often manage grades, deadlines, and school finances across separate tools — grades in one place, budgets in another, deadlines nowhere reliable.

Solution

Cross-platform student academic and finance tracker built with .NET MAUI and ASP.NET Core. It unifies terms, courses, assessments, credit-weighted GPA, notifications, and budget tracking behind a JWT-secured Azure API with offline SQLite storage.

Impact

.NET MAUI + ASP.NET Core student life app: GPA analytics, academic planning, and finance tracking with Azure-hosted JWT API — plus unit tests, wireframes, and architecture docs.

Tech stack