Justin Jonany
Machine Learning Research Engineer
About Me
Currently at Thema AI creating end-to-end pipelines, revenue models, and multi-agent systems that help investors discover their next great acquisition.
Education
University of Waterloo
BMath Data Science
Graduating June 2027
GPA
86.27%
Standing
Term Distinction (All Terms)
Clubs & Leadership
Basketball Intramurals, Volleyball Intramurals, Indonesian Student Association PR Lead, Board Game Designer
Scholarship
University of Waterloo President's Scholarship of Distinction (2023)
Experience
Machine Learning Research Engineer
Thema AI
- Engineered end-to-end data enrichment pipelines using SERP APIs, web scraping, and LLMs for 6M+ product and company data
- Lead development of AI-powered acquisition target ranker using multi-agent systems
- Performed topic modeling and dimensionality reduction (UMAP) to standardize 10M+ data points
- Developed statistical revenue prediction models with 95% confidence bounds 20% tighter than traditional providers
- Built concurrent company search engine that identifies 3x more relevant companies than OpenAI while being 4x faster and 10x cheaper
Data Science Intern
Ontario Securities Commission (OSC)
- Developed graph-based algorithm to identify 500+ suspicious trader groups among 20M+ individuals
- Engineered hybrid pipeline combining NLP, OCR, LLM, and RAG with 95% accuracy
- Led development of LLM evaluation frameworks based on academic papers
Remote Web Developer
PT Alto Sentosa
- Increased customer engagement by 30% in three months by leading design and implementation of a responsive 10+ page React web experience.
- Partnered with creative and project teams to iterate on UX flows, driving a noticeable rise in customer inquiries and conversions.
Featured Projects
Spotlight
Forward Looking Active Retrieval Generation (FLARE)
July 2024 – Oct 2024
- Developed advanced RAG architecture based on research paper, specializing in long-form text generation
- Reduced LLM hallucinations by 80% through statistical analyses
- Increased accuracy of traditional LLM PDF extraction from 91% to 97%
- Multi-input deep learning model for structuring scientific abstracts based on research paper
- 40% faster training on PubMed 200k RCT dataset with 2M sentences through optimization.
- 90% accuracy with modified Bi-LSTM with token, character, and positional embeddings.
Phylogenetic Analysis of COVID-19 Variants
Feb 2022
- Constructed phylogenetic trees from Southeast Asian SARS-CoV-2 genomes using NCBI GenBank datasets and MEGA X workflows.
- Applied the Kimura 2-parameter model with maximum likelihood inference and 1,000 bootstrap replications to validate evolutionary relationships.
- Mapped viral mutation patterns and geographic spread, revealing Vietnam's variant most similar to the Wuhan reference while Thailand diverged across eight months.
- Recognized with a Gold Award at LIVI 2022.
- Achieved a 94% F1-score on Indonesian Padang food Image classification.
- Boosted baseline accuracy by 4% after fine-tuning TensorFlow Hub's EfficientNetV2 feature extractor.
- Engineered data augmentation pipeline to stretch sixty image per class training data.
- Built a car sales forecasting system that blends classical ML models with custom neural network.
- Implemented forward/backward passes and gradient descent enitrely from sctach, without TensorFlow or PyTorch, to create a blending model.
- Achieved 7.6% MAPE—35% better than best Scikit-Learn models—using MLP/DNN architectures with ReLU/Leaky ReLU.
Current Focus
Where I'm investing deep work hours right now, from decentralized data integrity research to better ways of validating unsupervised models in production.
Blockchain
Currently diving into blockchain technology, decentralized systems, and how distributed ledgers plus consensus mechanisms can solve data integrity problems.
A big part studying smart contracts, Web3 infrastructure, and how decentralized rails could be integrated with data pipelines to make them more transparent and trustworthy.
Unsupervised ML Evaluation
The core challenge with unsupervised learning is the lack of ground truth, so I am researching creative out-of-the-box methods to evaluate them
This exploration grows out of insider trader detection at OSC, and to my current work in evaluating company search engine results.
The goal is to build statistically sound evaluation frameworks that make unsupervised models production-ready without eyeballing.
Coursework Snapshot
Core math, statistics, and computer science classes.
Fall 2022 · 1A
Term GPA 89.60%- CS 135 Designing Functional Programs
- MATH 135 Algebra for Honours Mathematics
- MATH 137 Calculus 1 for Honours Mathematics
Winter 2023 · 1B
Term GPA 86.40%- CS 136 Elementary Algorithm Design and Data Abstraction
- MATH 136 Linear Algebra 1 for Honours Mathematics
- MATH 138 Calculus 2 for Honours Mathematics
- STAT 230 Probability
Spring 2023 · 2A
Term GPA 85.00%- CS 246 Object-Oriented Software Development
- MATH 235 Linear Algebra 2 for Honours Mathematics
- MATH 237 Calculus 3 for Honours Mathematics
Winter 2024 · 2B
Term GPA 85.40%- CS 241 Foundations of Sequential Programs
- CS 245 Logic and Computation
- CS 251 Computer Organization and Design
- STAT 231 Statistics
Winter 2025 · 3A
Term GPA 84.75%- CS 240 Data Structures and Data Management
- CS 348 Introduction to Database Management
- MATH 239 Introduction to Combinatorics
- STAT 331 Applied Linear Models
Winter 2026 · 3B
Term GPA TBD- STAT 332 Sampling and Experimental Design
- STAT 341 Computational Statistics and Data Analysis
- CS 341 Algorithms
- STAT 333 Stochastic Processes 1
Beyond The Keyboard
When I am not ML-ing, you can usually find me steeping new teas, chasing endorphins at the gym or on long runs, and experimenting in the kitchen with bold flavors and recovery-friendly meals. I also love korean tv shows and alternative music!
Tea Enthusiast
My pantry is a rotating library of lapsang souchong, pu-erh bricks, gyokuro, genmaicha, smoked earl grey... (the list keeps going)
Gym, Running, and Hiking
Me and my friend hiking in Cold Springs, New York
Cooking
Weekends are for reverse-engineering restaurant dishes, replicating dishes from restaurants, and perfecting post-run recovery meals