I turn ambiguous needs into working systems.

I build learning tools, analyze telemetry, and turn data into software people can use.

Selected work

Midnight

An AI tutor grounded in a student’s own course material, with visual lessons and practice.

Rebuilt the learning loop from course ingestion through tutoring, quizzes, and flashcards. The current version is in user testing.

My role: Co-founder · led the rebuild and coordinated 25 contributors
Evidence: Authored 1,684 of 3,607 commits in the rebuild (about 47%). Earlier beta: 1,500 registrations.

Inspect Live product Case study Team case study

System architecture

STUDENT

PRODUCT UI Next.js · React · Radix UI · Tailwind

APPLICATION BACKEND server routes · API layer

Backend services

AUTH WorkOS

DATA PostgreSQL · Supabase

CACHE Redis · Upstash

AI ROUTING Vercel AI SDK

Background work, observability & deployment

API-triggered BACKGROUND JOBS Trigger.dev · notifications

App instrumentation MONITORING Sentry · PostHog

Build & deploy DELIVERY Turborepo · GitHub Actions · Vercel

OpenHouse Intelligence

A local AI assistant that turns an agent’s notes into CRM updates, with human approval before changes.

Built and pitched in one day. Top-8 finalist at the Dell × NVIDIA BuilderBase Hackathon in Seattle.

My role: Dashboard · integration · demo & pitch · team of 4
Evidence: The source includes a local SQLite CRM, a FastAPI backend, and an approval gate before proposed changes are saved.

Inspect Source

System architecture

Proof index

Telemetry, applied ML, and learning tools.

Additional project evidence
Project	Ownership	System	Maturity	Inspect
03 Motorsport Telemetry Analytics	Solo developer · Python analysis through deployed frontend	Static telemetry web app using real FastF1 data and a pace model evaluated on held-out stints 0.40s MAE on held-out Monza stints · 46% below a mean-prediction baseline	Live	Live demo Source
04 Wearable Health Telemetry ML	ML and backend · with Yash Piratla on the iOS app	Leakage-safe reproducible pipeline with fixed seeds, unit tests, cross-validation, and a model card 0.99 best macro-F1 on synthetic benchmarks, labeled for exactly what the benchmark can prove	Research · not a medical device	Source and case study
05 Snap-a-Mistake	Solo developer · Expo/React Native and Fastify	Three-stage transcription, image-fidelity, and independent diagnosis-verification pipeline 571 automated tests · disagreement is surfaced as suspect rather than certain	Open source	Source
06 Prometheus	IncludAI Neurodiversity Hackathon project	Learner-controlled grounded card creation, adaptive presentation, and FSRS scheduling Two learner-controlled AI loops: grounded creation and adaptive presentation	Live prototype	Live demo Source

03 Motorsport Telemetry Analytics Monza pace model: 0.40s MAE on held-out stints Live

Ownership: Solo developer · Python analysis through deployed frontend
System: Static telemetry web app using real FastF1 data and a pace model evaluated on held-out stints
Evidence: 0.40s MAE on held-out Monza stints · 46% below a mean-prediction baseline

Live demo Source

04 Wearable Health Telemetry ML Wearable-signal classification tested on synthetic data Research · not a medical device

Ownership: ML and backend · with Yash Piratla on the iOS app
System: Leakage-safe reproducible pipeline with fixed seeds, unit tests, cross-validation, and a model card
Evidence: 0.99 best macro-F1 on synthetic benchmarks, labeled for exactly what the benchmark can prove

Source and case study

05 Snap-a-Mistake Finds the first unsupported step in handwritten math Open source

Ownership: Solo developer · Expo/React Native and Fastify
System: Three-stage transcription, image-fidelity, and independent diagnosis-verification pipeline
Evidence: 571 automated tests · disagreement is surfaced as suspect rather than certain

Source

06 Prometheus Course-grounded flashcards that learners control Live prototype

Ownership: IncludAI Neurodiversity Hackathon project
System: Learner-controlled grounded card creation, adaptive presentation, and FSRS scheduling
Evidence: Two learner-controlled AI loops: grounded creation and adaptive presentation

Live demo Source

Experience

Product and engineering leadership, applied machine learning, and client delivery across education technology and automation.

Aug. 2021 to present

Co-Founder and Product/Engineering Lead, Midnight

I coordinate 25 contributors and build the learning loop for a course-grounded AI tutor, from course ingestion and tutoring to visual lessons, quizzes, flashcards, and AI routing.
2022 to 2024

Machine Learning Intern, Project Read

I worked with a five-person technical team on early-literacy tools, building and testing models that map generated phonemes to graphemes and contributing to their evaluation and refinement.
Aug. 2023 to Jul. 2026

Co-Founder and Technical Operations Lead, Praevius

I helped deliver web, voice, and automated acquisition systems through our AI automation agency, which served 12 small-business clients.

About

Builder and data scientist

I build data-driven tools and enjoy the work of making them useful to a real person, from the first model or interface through testing and iteration.

Education: B.S. in Data Science, Purdue University · Expected May 2029 · Sophomore standing through accepted transfer credit
Focus: Applied machine learning, telemetry and time-series data, product engineering
Based in: West Lafayette, Indiana
Outside work: Motorsport analytics, Muay Thai, and classical and jazz piano

Contact

Start a conversation

If you’re hiring an intern, mentoring an ambitious builder, or assembling a research team, I’d like to hear what you’re working on.

jmannana@purdue.edu

I turn ambiguous needs into working systems.

Featured project evidence

Co-Founder and Product/Engineering Lead, Midnight

Machine Learning Intern, Project Read

Co-Founder and Technical Operations Lead, Praevius

Start a conversation