Selected Work
Machine learning, civic technology and full-stack products. Source and live demos are linked wherever the repository is public; private work is described in full and available on request.
Screen 1 of 4: Planning a trip
LiveTTP Capstone
An AI travel planner that brings itineraries, visa checks, activities, packing lists and trip management into one app.
My contribution: Developed and secured 20+ REST API endpoints and contributed across the React client and the Express and PostgreSQL server of this team capstone.
Screen 1 of 5: Today
LiveFamily Planner
A family planner that puts the calendar, chores, meals, documents, notes and shared money behind one login.
My contribution: Built it solo: the schema, the row-level security policies that keep each family's data separate, the client, offline behavior and release.
Screen 1 of 2: Home page
LiveBMCC Hackathon
Helps CUNY students find the aid, benefits, scholarships and internships they may qualify for, ranked by fit.
My contribution: Built the deterministic eligibility engine that runs before any AI is consulted, the server-side AI workflows, and PostgreSQL persistence with authentication and rate limiting.
Six self-reported signals become a daily wellness snapshot, with every number traced back to the answer that produced it, and every claim the data could not support removed rather than reworded.
A low-risk general wellness tool, not a medical device. It does not diagnose, treat, monitor for crisis, or provide emergency services, and nothing entered into it is reviewed by a healthcare professional.
Music recommender that shows its working: every track carries a score, a confidence level, the reasons behind it, and the license it is served under. Built as the CodePath AI Engineering capstone.
Create playlists, add songs to them and play a thirty-second preview of each. It is a full PERN application built around a relational model rather than a flat list.
Structured résumé generation with live preview and browser-based PDF export. Entered data stays in the browser.
Merge, split, compress, rotate, watermark, sign, read and convert PDFs, entirely in your browser. Files are never uploaded to a server, because there is no server.
Built and shipped a sidewalk crack-detection service as the sole engineer. It covered the dataset pipeline, model training and evaluation, a FastAPI inference API and an interactive Streamlit dashboard.
Research prototype trained on 84 images and evaluated on 18. No cross-validation and no confidence intervals. Supports inspection prioritization; not a substitute for certified engineering assessment.
Two private study platforms I built for full-stack software engineering and interview preparation. Access is by invitation only, with a separate account for each person and no public sign-up.
A private, installable study system for full-stack interview preparation. It works offline once you are signed in, and each person's progress stays on their own account.
A private learning platform with its own server-side sign-in: structured curriculum, a code lab and a project archive, backed by a PostgreSQL database.
There is no sign-up. Requests go to me by email and I add each person myself. Or get in touch →
The code behind them is private; more of my work is on GitHub ↗
Featured Case Study
Problem
During my Machine Learning Internship at Walkspan ML, the model initially produced IoU = 0.0 on held-out images, predicting all background. CrackForest masks used label values of 1 and 2, but the preprocessing divided them by 255, making the ground-truth masks appear nearly empty and causing the model to learn nothing useful.
Action
Result
A second, subtler defect surfaced during that audit: the decoder recognized the label set only when it was exactly {1,2}, so 22 of 118 masks silently decoded to all-background. That left 18.6% of the ground truth reading as blank pavement while 78 tests kept passing. Correcting the decoder alone raised test IoU from 0.344 to 0.425 with no retraining, because the bug had been scoring the model's correct predictions as false positives. The released model scores 0.433 IoU and 0.595 Dice/F1 at threshold 0.85 on 18 held-out images, and the health-check fix cut worst-case API latency from 3.2 s to 1 ms. Shipped with 5 FastAPI endpoints, a locally verified Docker image, 141 passing tests, and a Streamlit research dashboard.
Evaluation Metrics
Stack
Availability
Available upon requestLet's Connect
I am actively looking for internships and early-career software engineering roles, full-stack with a strength in backend and data work.
