JM GitHub ↗

Johaan Mannanal

Data science for real systems.

I build data products for real systems: Midnight, an AI study platform rebuilt from an MVP that drew 1,500 sign-ups; a live F1 telemetry dashboard; and a reproducible ML pipeline for wearable heart-signal research. Purdue Data Science, Class of 2029.

View the work GitHub LinkedIn Résumé
Johaan Mannanal Incoming: West Lafayette, IN
SCROLL

About

Most student projects die as notebooks. Mine run. Verified numbers, stated limits, and work a stranger can open, test, and judge for themselves.

  • Education Data Science, Purdue University. Class of 2029, entering with sophomore standing
  • Focus Applied machine learning, telemetry and time-series data, product engineering
  • Based in Incoming to West Lafayette, Indiana, August 2026
  • Open to Internships, research, and racing or robotics project teams

The work

Built. Tested. Shipped.

One product I help run, and two projects you can open right now. Every number is either reproducible from a repo or labeled as reported.

Flagship · In user testing

Midnight

The AI study platform I build every day

Lead Engineer · Product Co-lead

Midnight turns scattered class inputs into finished work: AI grading, quiz generation, spaced-repetition flashcards, a tutoring workspace, and Canvas-synced tasks. Our first beta drew 1,500 registrations and 10M+ social-media views; the rebuild at midapp.me applies everything we learned from it and is now in user testing. I led that rebuild and wrote nearly half the code across a 10+ person team.

47% of 3,600+ commits, top contributor
5 AI-assisted tools shipped
10+ contributors coordinated
  • TypeScript
  • Next.js
  • Supabase
  • WorkOS
  • Trigger.dev
  • Multi-provider AI
Live product Case study
Live Motorsport Telemetry Analytics preview
02

Motorsport Telemetry Analytics

Solo Developer · Python analysis to deployed front end

Pick two drivers from a Grand Prix and see exactly where the lap time lives: braking points, sector splits, tyre life, and a pace model evaluated on whole held-out stints. Real FastF1 telemetry, drivers in their real team colours, rebuilt as a static web app that loads in a blink.

0.40s MAE on held-out Monza stints, 46% below a mean-prediction baseline
  • Python
  • TypeScript
  • FastF1
  • scikit-learn
  • Plotly
Live demo Source
Wearable Health Telemetry ML
ModelsSVM ensemble · Gradient Boosting · Neural net
RigorLeakage-safe pipelines · fixed seeds · 5-fold CV
HonestyModel card · verified metrics · stated limits
03

Wearable Health Telemetry ML

Researcher · ML Developer

Four models, one honest question: what can classifiers really learn from wearable heart signals? A fully reproducible pipeline with fixed seeds, unit tests, and a model card that says out loud what synthetic data can and cannot prove. Research, not a medical device.

0.99 best macro-F1 on the synthetic benchmarks, labelled for exactly what it is
  • Python
  • scikit-learn
  • NumPy
  • pandas
  • pytest
Source & case study

Stack

What I reach for

Chosen for work that has to hold up, not just demo well.

Languages

  • Python
  • TypeScript
  • JavaScript
  • SQL
  • Java

ML & Data

  • scikit-learn
  • pandas
  • NumPy
  • Plotly
  • Matplotlib

Web & App

  • Next.js
  • React
  • Node.js
  • Astro
  • Streamlit

Infra & Tools

  • Git
  • Docker
  • Vercel
  • Supabase
  • pytest
  • Linux

Contact

Let's build
something real.

Open to internships, research labs, and Purdue project teams. Email gets the fastest answer; everything else gets there eventually.

johaan.joseph@gmail.com