Seeking a Summer 2027 co-op in software engineering, ML / AI engineering, data science, or quant developmentComputer Science Co-op Student | Software Engineering | Full-Stack | ML
I build software, run experiments with machine learning, and want to understand the math underneath.
GitHub Resume (PDF) jeeshaanbakshi@gmail.com
Ask my portfolio
Search everything here by meaning. Try one of these:
Study groups for Waterloo students taking the same courses.
Put security where the data lives. RLS checks every query inside Postgres, so the UI never has to be trusted.
ProblemClient-side permission checks can be bypassed.
FixMoved authorization into PostgreSQL RLS policies keyed on auth.uid().
Find the passages that answer a question, by meaning, not keywords.
With normalized embeddings, cosine similarity is a dot product. Retrieval is linear algebra.
ProblemBig chunks match vaguely; small chunks lose context.
FixTuned chunk size and overlap to balance the two.
This site. A map of my work and search that runs in your browser.
One typed data file feeds the pages, the search index and the resume, so they can't drift apart.
ProblemML search with no backend could slow the page for everyone.
FixThe model loads only on the first query.
A phone-controlled car. My circuit, my C++ firmware.
Change one thing at a time. It finds a hardware fault faster than swapping parts.
ProblemControl kept dropping from hardware and connection faults.
FixTested component by component until control was reliable.
auth.uid() = owner_id
The check a row-level security policy runs on every query (design)
cos θ = a·b / ‖a‖ ‖b‖
Cosine similarity between 384-dimensional MiniLM embeddings
z = Wᵀ(x − μ)
PCA: each 384-d passage embedding projected onto two principal components
Study groups for Waterloo students taking the same courses.
ProblemClient-side permission checks can be bypassed.
FixMoved authorization into PostgreSQL RLS policies keyed on auth.uid().
Find the passages that answer a question, by meaning, not keywords.
ProblemBig chunks match vaguely; small chunks lose context.
FixTuned chunk size and overlap to balance the two.
This site. A map of my work and search that runs in your browser.
ProblemML search with no backend could slow the page for everyone.
FixThe model loads only on the first query.
A phone-controlled car. My circuit, my C++ firmware.
ProblemControl kept dropping from hardware and connection faults.
FixTested component by component until control was reliable.
01
I get the smallest version working end to end first.
RAG Retrieval System02
When something breaks, I change one thing at a time.
Bluetooth car03
Then I tune it. Most choices are trade-offs, like chunk size vs. context.
RAG Retrieval System04
I make the safe option the default, like permissions inside the database.
CourseMateAbout
I like understanding systems from the inside out.
Web apps, ML pipelines, data work and Arduino hardware.
I want to know why something works, where it breaks, and how to make it better.
Half my time goes into building it. The other half goes into making it right.
I like understanding systems from the inside out.
Web apps, ML pipelines, data work and Arduino hardware.
I want to know why something works, where it breaks, and how to make it better.
Half my time goes into building it. The other half goes into making it right.
Off the keyboard
robotics · math problem-solving · tennis · table tennis · badminton · cycling
Toolkit — with where I've used it
Data Science Intern
UAE · 2025
Savvy Data Cloud Consulting
University of Waterloo
Waterloo, ON, Canada · Sept 2026 – Expected 2031
Honours Bachelor of Computer Science, Co-op
Relevant coursework
2024
AMO Gold Medal
American Mathematics Olympiad
✦
All-India Rank 64, School Rank 1
SOF International Mathematics Olympiad
2025
SASMO Bronze Medal
Singapore and Asian Schools Math Olympiad
Recruiting for a Summer 2027 co-op? Send a note here or email jeeshaanbakshi@gmail.com.
Spotted a bug on this site? .