P. Kanisius Bagaskara

Machine Learning Engineer (Student)

Jakarta, Indonesia

0%

Preparing Pit Lane...

KANISIUS
BAGASKARA
ML EngineerJakarta, Indonesia
Scroll
Selected Work

Featured Projects

Production-ready ML systems with real-world impact. Each project includes problem analysis, technical implementation, and measurable results.

01

F1 Dashboard Analytics

Real-time telemetry & ML-powered strategy predictions

Live DemoCodeCase Study
The Problem

F1 fans struggle to understand race strategies. Existing tools lack real-time ML predictions for tire degradation and optimal pit stops.

My Solution
  • →Interactive telemetry visualization using FastF1 API
  • →XGBoost model predicting optimal pit windows (92% accuracy)
  • →Driver performance comparison across 23 race tracks
  • →Weather integration affecting strategy recommendations
Impact
92%
Model Accuracy
<50ms
API Response
23
Race Tracks
Challenges Solved
  • •Handled missing telemetry data with interpolation algorithms
  • •Optimized API fetching within rate limits
  • •Built mobile-responsive UI for live race viewing
PythonStreamlitFastF1XGBoostFastAPI
02

Fraud Detection

Binary classification for financial transaction security

Live DemoCode
The Problem

Financial fraud detection requires handling imbalanced datasets and complex transaction patterns.

My Solution
  • →XGBoost classifier with comprehensive feature engineering
  • →Transaction pattern analysis & anomaly detection
  • →Handled class imbalance with SMOTE
  • →ROC-AUC optimization for fraud detection
Impact
0.94
ROC-AUC
590k+
Dataset
400+
Features
Challenges Solved
  • •Extreme class imbalance (3.5% fraud cases)
  • •Feature engineering from anonymized data
  • •Time-based validation to prevent leakage
XGBoostPandasScikit-learnSeabornMatplotlib
03

Deep Dive into Machine Learning

Machine Learning algorithms implemented from scratch (CS229)

Code
The Problem

Understanding the inner workings of core algorithms to build robust and optimized AI systems.

My Solution
  • →Implemented classic ML algorithms purely in NumPy
  • →Built custom optimization engines (Gradient Descent, Adam etc.)
  • →Mathematical derivations for backpropagation and loss functions
  • →Compared performance against enterprise libraries (Scikit-Learn)
Impact
8+
Algorithms
≈ Parity
vs Scikit-Learn
100%
From Scratch
Challenges Solved
  • •Deriving complex calculus for backward passes
  • •Optimizing matrix multiplications without frameworks
  • •Handling numerical instability (vanishing gradients)
PythonNumPyMatplotlibJupyterMath
04

Spotify Downloader

Real WAV/FLAC/MP3 from a playlist link — no account, verified codec

Code
The Problem

People want their own offline playlists in edit-friendly WAV. Most tools just wrap an MP3 in a .wav container, and lossless routes need a Spotify account or violate its ToS.

My Solution
  • →Spotify link used as metadata source only — no account, no login
  • →Provider ladder + scoring engine picks the best match per track
  • →In-app PCM decode and mux — WAV verified as pcm_s16le before saving
  • →One TypeScript codebase serving Next.js web and Electron desktop
Impact
WAV·FLAC·MP3
Formats
pcm_s16le
Codec Verified
MVP E2E
Milestone
Challenges Solved
  • •Rejecting fake WAV — an MP3 stream wearing a .wav container
  • •Matching tracks without the Spotify API (normalized title/artist, duration tolerance)
  • •Multi-provider quality ladder instead of one flaky source
TypeScriptNext.jsNode.jsElectron
05

Progressive Live 3D ReconstructionPrivate · WIP

Phone camera in, 3D model out — reconstructed live on the laptop, no cloud

Code
The Problem

Capture tools split in two: offline photogrammetry makes you wait minutes to hours, on-device scanning locks quality to phone hardware. Can the phone stay dumb and the laptop do the work, so the first shape lands in seconds?

My Solution
  • →Phone thin client: ARCore pose, frame quality gate, JPEG over WebSocket (WiFi or USB)
  • →One mutable reconstruction state, single worker — viewers read snapshots
  • →Live 3D Gaussian Splatting via gsplat, measured 1.59 ms/frame on an RTX 2050 4 GB
  • →Exports .splat plus a self-contained Three.js viewer
Impact
1.59ms
Live Render
5MiB
VRAM Peak
163
Tests
Challenges Solved
  • •4 GB VRAM is the binding constraint — anything that doesn't fit gets sequenced or dropped
  • •Dense stage stays gated: no COLMAP binary to prove it, so the exporter fails loudly instead of writing a bad mesh
  • •Scanning below ARCore's documented 0.5–5 m band — pose quality is still an open question
PythonFluttergsplatOpen3DWebSocketARCore
View More on GitHub
Skills

Technical Stack

Tools and technologies I use to build, experiment, and learn. I focus on data, machine learning, and building real-world applications.

Always learning

This stack evolves as I explore new technologies and work on new projects.

Programming

Languages I use for data analysis, machine learning, and building applications.

Python
Primary language
SQL
Querying & modeling
TypeScript
Web & tooling

Data & Machine Learning

Libraries for data processing, analysis, and machine learning.

Pandas
Data manipulation
NumPy
Numerical computing
Scikit-Learn
ML algorithms
PyTorch
Deep learning
Matplotlib
Visualization
Seaborn
Statistical plots
Plotly
Interactive charts

AI Systems & Tools

Tools for building and experimenting with modern AI systems.

Claude AI
LLM integration
Codex / OpenAI
Code generation
Jupyter
Experimentation
Google Colab
Cloud notebooks

Engineering & Deployment

Tools for building, deploying, and maintaining applications and data pipelines.

FastAPI
Backend development
Docker
Containerization
GitHub Actions
CI/CD
GitLab CI
DevOps
Streamlit
Data apps

Other Tools

Tools I use for productivity and development.

Git
Version control
VS Code
Code editor
Linux
WSL2 dev env
Windows
Daily use
Exploring · Building · ImprovingCurrently a student · Open to internship opportunities
Track Record

By The Numbers

Quantifiable impact through code, education, and continuous learning.

8+
Repositories
Top 100
Rising Star Award
23
Tech Stack
5
Major Projects

Technical Impact

4+Advanced ML Projects

Focusing on predictive analytics

10M+Telemetry Frames

Processed in F1 analytics

92%Pit-Window Accuracy

F1 strategy model

Education & Teaching

Top 100Rising Star Award

Fully funded graduation invite

5+AI Workshops

Conducted for students

100+Students Trained

In AI and data science

Recognition

StanfordML Specialization

3 courses completed

GoogleGemini Certified

Educator & Student

IBMData Science

Professional certificate

Continuous Learning

Certifications

Every certificate here was earned to solve a real problem — not just to collect badges. Stanford for the math, Google for the tools, IBM for the foundations.

Jan 2026
Click individual courses to verify
IBM

Python for Data Science

“The foundation. Where it all started.”

Aug 2025
Verify
Google for Education

Gemini Certified Educator

“Bridging AI tools with real teaching impact”

Sep 2025
Verify
About Me

Turning Data Into Racing Insights

Machine Learning Engineer focused on building production-ready ML systems and real-time analytics. Shipped an F1 analytics dashboard with ML pit-window predictions across all 23 races, and now building progressive live 3D reconstruction — phone camera in, 3D model out.

I've conducted AI workshops for 100+ students, making complex AI concepts accessible through hands-on learning.

My approach combines rigorous data research analysis with First Principles Thinking —stripping complex problems down to their fundamental truths before building scalable ML architectures.

Kanisius Bagaskara

Official Representation

Rising Star Top 100
Stanford ML Specialization
8+ Open Source Projects
100+ Students Trained

Core Expertise

Claude AI, OpenAI Codex, LLM Prompting

AI Integration

First Principles Thinking, Statistical Modeling

Data Research

Docker, CI/CD, Model Deployment

MLOps

FastF1 API, Version Control, Pipelines

Data Engineering

AI Workshops, Mentoring

Leadership

Open for Opportunities

ML Engineering internships & freelance projects

Random Facts

How I Got Here

01

The Iron Man Thing

Honestly? I got into Data Science because of Iron Man. Watching JARVIS talk, analyze data in real-time, help Tony make decisions—I thought, “Damn, I wanna build something like that.” Started learning Python, fell into ML, and now I'm stuck here (in a good way).

02

From Bedroom to Stage

Used to be just some kid coding alone in my room. I was the guy who got nervous talking to 5 people. Now I can teach workshops with 100+ people. Didn't see that coming, but turns out teaching is fun—it actually makes me understand stuff better.

03

The Stubbornness

I never aim to be the best. I just know that when I have a target, I chase it until I get it. Like F1—it's not about being perfect, it's about improving every single lap. Code broke? Fix it. Algorithm failed? Try again. Just keep moving forward.

I'm not perfect, but when I have a goal,
I DON'T STOP UNTIL I GET IT.

Too stubborn to quit
Always LearningToo Stubborn To QuitF1 > Football
Message from Kanisius

“Failure is an option here. If things are not failing, you are not innovating enough.”

Live

Currently

Recent Win

Completed Stanford ML Specialization (3 courses)

Last updated: February 2026

Get In Touch

Let's Build Something

Currently open to opportunities. Usually respond within 24 hours.

Open For

ML Engineering internships
Freelance data science projects
F1 analytics collaboration
Guest speaking at tech events
Start a ConversationDownload CV
Usually respond within 24 hours
Email
kanisius@novar.me
LinkedIn
linkedin.com/in/kanisiusbagas1212
GitHub
github.com/knsiuss
Instagram
instagram.com/p_knsss