I build the model, and the product around it.

A medication tracker for iPhone. My dad started maintenance meds early in life, and it’s my mom who keeps him on track. Not everyone has someone like that. Deni is meant to be that reminder for anyone who doesn’t, or for anyone who simply wants something dependable to lean on. Deni organizes your meds like they’re all in one cabinet, so you always know what to take and when. A journal tracks how you’re feeling alongside your doses, and Insights sums up your week in a one-page report you can share with your doctor.

• React Native
• Expo
• TypeScript
• SQLite

A human-feedback platform for AI teams: contributors compare two AI answers, say why one is better, and are paid in USDC on Stellar as soon as the answer is accepted. Gold tasks, rate limits, and agreement checks keep the rankings honest. Payouts are co-signed from a multisig account, and after most balances sat stuck below the old withdrawal minimum, the Withdraw button went away: every accepted answer now pays the wallet directly.

Instawards grant

$5,000

keys co-sign every payout, enforced in CI

2 of 3

• Next.js
• TypeScript
• PostgreSQL
• Prisma
• Stellar SDK
• Redis
• Railway

Classifies exercises and grades form in real time on a Raspberry Pi 5. BlazePose landmarks feed a dual-head LSTM trained on 451,638 augmented 30-frame windows. After a session, a RAG chat explains what to fix: retrieval runs on the Pi with an ONNX encoder, so PyTorch never ships to the device.

exercise classification, 9 movements

97.15%

live inference on a Raspberry Pi 5

25–30 fps

• MediaPipe
• LSTM
• TFLite
• ONNX
• Groq
• Raspberry Pi

Answers clinical questions from 100+ pages of the 2024 ESC atrial fibrillation guidelines. MedCPT embeddings, FAISS search, and BGE reranking ground three quantized models (Llama 3, Phi-3, Qwen3), tested on 20 clinical queries with a case-by-case look at every hallucination.

BERTScore F1

0.835

response latency, quantized

~0.7 s

• Llama 3
• Phi-3
• Qwen3
• FAISS
• MedCPT

Finds cysts, stones, and tumors in axial and coronal CT slices with YOLOv12, trained on 14,761 curated and augmented scans. Presented at IEEE ICIPCN 2026 at Kathmandu University, Nepal.

mAP@0.5, axial slices

0.946

mAP@0.5, coronal slices

0.885

• YOLOv12
• Python
• Streamlit

Reads seven facial expressions from a webcam and answers each one with GSAP-driven visual feedback. The face-api.js models run entirely in the browser.

• React
• face-api.js
• GSAP

A dashboard for running apartment properties: tenants, units, and live Firebase data with charts. Demo login: test@test.com / test123.

• React
• Firebase
• SCSS

Keeps a pet’s health records, vet appointments, and daily care routines in one place. Built with React and TypeScript on an Appwrite backend.

• React
• TypeScript
• Appwrite