Projects

First complete PCB: KiCad schematic through two-layer layout and Gerbers for an Arduino Nano, HC-SR04, and status LED.

Motivation: I built this to learn the real board house pipeline, not to invent a fancy circuit. Custom HC-SR04 footprint from measured dimensions, two-layer routing with a ground plane, DRC, then Gerber and Excellon export for PCBWay. Breadboards were familiar. Owning symbols, copper, mask, and fab files was the point.

Challenge: Stock connector footprints did not match the real ultrasonic module.

Built a deployed personal media archive that separates perceived film quality from personal impact, captures short post-watch reflections, and organizes them into a searchable theme-based knowledge base.

Motivation: I kept remembering that a movie was an 8/10 and forgetting why it stayed with me. Existing trackers are excellent at logging titles and scores; they lose scenes, questions, and takeaways. I designed Afterthought as a Next.js App Router product with dual-axis ratings (Overall vs Personal Impact), a hard 2–3 minute capture constraint centered on one prompt (“What stuck with you?”), browser voice notes with transcription and AI organization, TMDB-backed media identity, and a persistence boundary that runs on a seeded Zustand demo store or Supabase auth/Postgres/storage. The hard problem was not adding fields. A multi-section journal felt like homework, so I forced capture and organization apart: the user writes the thought; AI may suggest themes afterward, never overwrite the original.

Challenge: A richer schema (lessons, scenes, quotes, techniques) made post-movie capture feel like a questionnaire.

Lending Club default model for CIS 5450: clean the data, kill leakage, then rank risk under a messy class balance.

Motivation: Course project on ~1M Lending Club loans. I filtered to paid vs charged-off, threw out post-outcome cheat columns like recoveries and total_pymnt, and only fitted transforms after the train/test split. Models went from logistic regression up to tuned XGBoost around 0.7175 AUC. Tuning barely moved the needle once the columns were honest. Accuracy was a trap with an 80/20 split, so I cared about AUC, PR, and the fact that a 0.5 threshold is not a business decision.

Challenge: Some of the strongest-looking columns only exist after default.

Research outreach helper: scrape lab pages, draft with a local LLM option, and never send until I review.

Motivation: Cold email prep was eating time, so I automated the boring parts and refused to automate the dangerous one. LabReach finds faculty pages, pulls context, drafts with Ollama when I want text to stay local, and writes a review artifact before Gmail can do anything. JS-heavy department sites break naive scrapers. That is fine. Guessing an email and blasting it is not.

Challenge: JS-heavy faculty pages break static scrapers.

Dance practice tool: scrub a waveform, get BPM for that section, keep rehearsing.

Motivation: I kept losing time hunting song sections and tempos, so I built the tool I wanted. Pick a region on the waveform, send it for tempo analysis, get numbers back without breaking rehearsal flow. Started in Colab, ended as a deployed Next.js app with Python on the backend. When the selection edges are wrong, the BPM looks wrong, and dancers blame the model. Sync mattered more than fancy DSP.

Challenge: Selection boundaries must match what the analyzer hears.

Handheld Tetris on Arduino Nano + ST7735: game loop, SPI draw, buttons, and sound all on-device.

Motivation: I wanted a real embedded game loop, not a simulation. Everything runs on the Nano: gravity, collision, scoring, SPI redraw, piezo cues. Half the 'bugs' were solder joints on the display. I staged controls on a two-button harness first so I was not debugging mechanics and firmware at the same time.

Challenge: Intermittent SPI looked like sprite corruption.

CV dance coach: compare your form to a reference clip and get cues that actually mean something.

Motivation: I dance, so I care when feedback is useless. Bhangra Coach uploads a clip, runs MediaPipe landmarks, aligns against a reference when tempos differ, then surfaces deltas in a Next.js UI with FastAPI and Supabase behind it. Raw joint error is not coaching. Alignment and wording are the product right now, and both still need work.

Challenge: Tempo differences break naive frame-wise compare.

Campus dining social app from Penn SPARK: auth, profiles, and a small student loop that actually shipped.

Motivation: SPARK project with a real deadline. We built Penn Plates on Next.js and Supabase so students could coordinate around dining without a bloated feature list. Auth and RLS had to be right. Scope cuts hurt, but they are why we had a demo instead of half-finished screens.

Challenge: Ambition versus timeline.

Living engineering portfolio that turns projects into structured technical case studies instead of a static screenshot gallery.

Motivation: I built this site so recruiters and engineers can evaluate my work the way I want it evaluated: systems thinking, design decisions, iterations, evidence, and reflection. Hardware, software, research, embedded, ML, CAD, and product work all land in one visual and narrative system. Typed project data drives cards, routes, heroes, and media. Case studies attach by slug. Featured work and supporting builds get different homepage weight so experiments stay archived without competing with flagship projects. Adding a project should feel like publishing into a system, not redesigning a page.

Challenge: Earlier pages showed outcomes without process, failures, or tradeoffs.