Skip to content
Paul Mandap

Work

Selected projects

Three of these have full write-ups — the pipeline I built at Stratpoint, and two systems I built on my own hardware because I wanted to find out where they break.

01Case studies

Personal project · live, in use

Study

An installable study app that turns a student's own notes into flashcards and quizzes, shows the page every answer came from, and grades written answers against a rubric — built for five real users on free tiers, with every product decision written down before it was built.

447 tests passing
13.3k lines of TypeScript
14 logged product decisions
TypeScriptReact Native WebExpo RouterSupabasePostgreSQL+4

Stratpoint Technologies · Data Engineering Internship

Movie Analytics Warehouse

An end-to-end ELT pipeline that turns three messy source files into a 20-model dimensional star schema, built for exactly one reader: a specialty-insurance underwriter deciding what premium to quote on a film.

20 dbt models · 3 layers
1m22s full warehouse rebuild
104 dbt tests
Apache Airflow 3.3dbt 1.12PySpark 3.5PostgreSQL 16Docker+2

Personal project

Personal AI OS

A model-agnostic personal agent system running entirely on a consumer GPU — no cloud provider anywhere in the codebase — with a purpose-built evaluation harness, a protected holdout split, and 69 architecture decision records including the ones that killed my own hypotheses.

1,058 tests · 1,042 offline
69 decision records
85 eval cases · 14 suites
Python 3.11OllamaSQLitepydantichttpx+2

Personal project · shipped to external testers

Nod

A Windows overlay that listens for a wake word, answers out loud in a locally-synthesised Philippine-English voice, and keeps a rolling summary of the meeting you're in — with microphone and speaker audio kept structurally incapable of reaching the same code path.

98.4% intent accuracy
606ms p50 classification
2–4s end to end, CPU only
Python 3.10PyQt6faster-whisperPiper TTSOllama+3

02Also built

Full-stack project

AI-Powered Volleyball Gesture System

A gesture recognition backend built on YOLO, MediaPipe and Keras, wired through FastAPI and WebSockets into a Next.js frontend so a referee's hand signal updates every connected scoreboard instantly.

YOLOMediaPipeKerasFastAPIWebSockets+1

Machine learning project

NLP Resume Classifier

An end-to-end pipeline using spaCy NER and an SVM classifier that parses resumes into structured records across 25 job domains, at 80% accuracy and a 0.77 F1-score.

spaCyscikit-learnSVMPython

Undergraduate thesis

Autonomous RC Car Navigation

Undergraduate thesis: a real-time YOLOv8n and ROS2 obstacle-detection pipeline with Kalman filtering and ORB-SLAM3, achieving 0.1m distance error at 10 FPS on constrained hardware.

YOLOv8nROS2ORB-SLAM3Kalman filterRaspberry Pi 5