UofT CS + Statistics · Class of 2025
Vinayak
Maharaj
Software Engineer — AI Systems
- Built end-to-end anomaly detection pipeline processing 100k+ log events weekly, cutting manual analysis time by 70% at Maritime Financial Group
- Shipped Inbox Copilot and Prepwise as production SaaS products with Stripe billing and real paying users
- Active research with TTLab on where zero-shot LLM reasoning breaks down in long-horizon decision environments
Currently
Open to SWE roles
Toronto, ON · Available now
100k+
Events / Week
70%
Analysis Time Saved
100+
Interview Sessions
16+
GitHub Repos
Focus Areas
Impact
Numbers that mean something
Every number comes from a real project, not a resume estimate.
100k+
Events / Week
Firewall and VPN log events through the Maritime pipeline weekly, reducing manual analysis time by 70%
1,000+
Emails / Month
Processed through Inbox Copilot, my own inbox fully indexed and queryable
100+
Interview Sessions
AI mock interviews conducted on Prepwise with real users across multiple roles
16+
GitHub Repos
Across SaaS products, ML systems, research projects, and backend engineering
Work
Things I have built
Across SaaS, ML systems, and personal projects. 16 repos on GitHub.
AI SAASFEATURED
Inbox Copilot
Problem
Email overload kills focus. Important threads get buried.
Solution
RAG pipeline over your entire inbox via Pinecone vector search. LangChain drafts replies with full context.
Outcome
Stripe billing, real paying users, Aurinko webhooks for live sync.
Full-stack AI email client. Indexes your inbox, surfaces relevant threads, drafts replies. Aurinko for real-time webhooks, tRPC for type-safe APIs, Prisma for data modeling, Stripe for subscriptions.
Next.jsLangChainPineconeAurinkoStripetRPCPrismaPostgreSQL
AI
AI Developer Collaboration Platform
RAG over GitHub repos. Ingest codebases into pgvector, query with natural language. AssemblyAI for voice-to-code, Stripe credits system, Docker, Fly.io.
Next.jspgvectorAssemblyAIStripeDockerFly.io
SAAS
Prepwise
AI mock interview platform. Vapi voice agent runs the conversation, Google Gemini scores answers and generates feedback. 100+ sessions conducted.
VapiGoogle GeminiFirebaseNext.jsTypeScript
ML
F1 Lap Time Predictor
LSTM trained on real Monaco GP 2023 telemetry via FastF1. Predicts next lap from previous 10. SHAP explainability, XGBoost baseline, interactive Gradio app.
PyTorchFastF1XGBoostSHAPGradioPython
FINANCE
Algorithmic Trading Backtest
Twitter engagement ratio strategy backtested against NASDAQ. Portfolio outperformed benchmark across 2022 by up to 25%. Signal generation, portfolio simulation, performance attribution.
PythonPandasNumPyMatplotlibyFinance
RESEARCH
NBA Hot Hand Bayesian Analysis
Compared pooled, unpooled, and hierarchical Bayesian priors on NBA shooting patterns using PyMC. Tested the hot hand fallacy statistically. Presented to the UofT Statistics department head.
PyMCPythonBayesian InferenceHierarchical ModelsStatistics
BACKEND
CommerceCore
Production-style e-commerce REST API in Java. JWT authentication, role-based access control, full cart/order/product management, clean layered architecture across Controller, Service, Repository, and DTO layers.
JavaSpring BootSpring SecurityJWTJPAMySQLMaven
FINANCE
Finanseer Dashboard
Finance dashboard with revenue/expense tracking, linear regression forecasting, and Redux Toolkit Query for state management. Built during internship at Amnevar Ltd.
ReactRedux Toolkit QueryTypeScriptRechartsNode.js
Skills
What I work with
Core Proficient Familiar
AI / LLM
Languages
Frontend
Backend
Data & Storage
Infrastructure
Experience
Where I have worked
Maritime Financial Group
- Built end-to-end firewall and VPN anomaly detection pipeline processing 100k+ log events weekly, reducing manual analysis time by 70% through automated risk scoring
- Hybrid detection system combining rules, Isolation Forest, and autoencoder over 5-minute windows to catch login bursts and multi-region access patterns with explainable scoring
- Next.js dashboard with grouped alerts, drill-down views, and trend analytics, cutting average analyst investigation time per incident by 40%
PythonPyTorchIsolation ForestAutoencoderNext.js
Intelligent Adaptive Interventions Lab, UofT
- Built interactive data visualization components (bar, pie, scatter) using Recharts on Next.js for a UofT research lab's adaptive experimentation platform
- Plugin-based visualization architecture with dynamic frontend loading and Flask backend analysis endpoints, designed for extensibility across future contributors
Next.jsRechartsFlaskPython
Amnevar Ltd.
- Replaced manual paper-based bookkeeping with a full-stack finance dashboard (React, TypeScript, Node.js, MongoDB), enabling real-time KPI tracking for the first time
- Linear regression revenue forecasting model and Redux Toolkit Query for state management, enabling next-year projections from historical data
ReactRedux Toolkit QueryTypeScriptNode.jsMongoDB
Education
University of Toronto · Trinity College · Mathematics Minor
- Bayesian analysis research presented to Statistics department head
- Specialization spanning ML, stats, algorithms, and software engineering
Technical Courses
CSC311H1
Machine Learning
Supervised/unsupervised learning, neural nets, SVMs, Bayesian classifiers
STA365H1
Applied Bayesian Statistics
Bayesian inference, MCMC, hierarchical models, posterior computation
STA457H1
Time Series Analysis
ARIMA, spectral analysis, forecasting — foundation for the F1 predictor
CSC373H1
Algorithm Design
Greedy algorithms, dynamic programming, network flow, NP-completeness
CSC263H1
Data Structures
Trees, hashing, priority queues, graphs, amortized analysis
CSC301H1
Software Engineering
Agile, testing, architecture, team delivery on production software
STA410H1
Statistical Computing
Numerical methods, simulation, optimization in statistical workflows
CSC207H1
Software Design
Design patterns, OOP principles, testing, UML
ENT200H1
Innovation & Entrepreneurship
Feasibility analysis, business model canvas. Designed a UofT study-space recommendation platform (surveyed 26 CS students, full market analysis) and a VR live sports events platform with revenue modeling.
Certifications
IBM + CourseraOct 2025
RAG and Agentic AI
Professional Certificate · 8 courses
LangChain, LangGraph, CrewAI, AutoGen, vector databases, multi-agent architectures.
AWS + DeepLearning.AIJun 2025
Generative AI with LLMs
Course Certificate
LLM training, fine-tuning, RLHF, deployment at scale.
DeepLearning.AI + StanfordJan 2024
Machine Learning Specialization
Specialization · 3 courses · Andrew Ng
Supervised/unsupervised learning, neural networks, decision trees, recommender systems, RL.
Now
Outside the editor
The stuff that does not go on a resume.
Currently Reading
Football
Barcelona fan for as long as I can remember. From Ronaldinho to MSN to watching Lamine Yamal do things at 17 that should not be physically possible. Watching Messi, Suarez and Neymar together as one attack is probably the greatest football I will ever see.
Man City came when Aguero was at his peak. When Guardiola signed I knew something special was coming. The 2023 treble playing some of the most beautiful football in Premier League history cemented everything.
Man City came when Aguero was at his peak. When Guardiola signed I knew something special was coming. The 2023 treble playing some of the most beautiful football in Premier League history cemented everything.
Cricket + Basketball
From Trinidad. Home of Brian Lara. Cricket is not something you choose in the Caribbean, it is part of who you are. Rally round the West Indies.
Kohli will always be my favourite. Dropped from the national team, written off, came back as undisputedly the greatest batter to ever play the game. That response to doubt stays with me.
Golden State Warriors since the Curry dynasty started. Steph was told his body was not built for the NBA. He became the greatest shooter the sport will ever see and won the only unanimous MVP in history. In the 90s everyone wanted to fly like Jordan. When I play basketball, I want to shoot like Curry.
Kohli will always be my favourite. Dropped from the national team, written off, came back as undisputedly the greatest batter to ever play the game. That response to doubt stays with me.
Golden State Warriors since the Curry dynasty started. Steph was told his body was not built for the NBA. He became the greatest shooter the sport will ever see and won the only unanimous MVP in history. In the 90s everyone wanted to fly like Jordan. When I play basketball, I want to shoot like Curry.
F1
Drive to Survive is where it started, like most people. What kept me was Max Verstappen. The ability to completely block out the noise and politics and just perform at an insane level while being genuinely the funniest person on the grid is something I take notes from. Hoping he breaks Schumacher's record.
The Monaco 2023 GP data through FastF1 has telemetry across all 78 laps. Throttle position, DRS, speed at every mini-sector, tyre compound. The LSTM needs a 10-lap window to predict the next lap. Tyre deg and sector consistency matter more than raw top speed, which is exactly how the best Monaco drivers think about the race.
The Monaco 2023 GP data through FastF1 has telemetry across all 78 laps. Throttle position, DRS, speed at every mini-sector, tyre compound. The LSTM needs a 10-lap window to predict the next lap. Tyre deg and sector consistency matter more than raw top speed, which is exactly how the best Monaco drivers think about the race.
Active Research
Can an LLM Play Balatro Without Ever Practicing?
Active research with Prof. Patrick Hosein at TTLab. The question: serialize a full Balatro game state into text, feed it to an LLM with zero training, and compare its decisions against the heuristic bot that ships with the Balatrobot API.
The contribution is not whether it wins. It is diagnosing exactly where zero-shot LLM reasoning breaks down in a long-horizon, sparse-reward environment. Balatro is a better testbed than the roguelikes used in prior work because every decision is discrete and interpretable — you can see where the reasoning fails, not just that it does.
The contribution is not whether it wins. It is diagnosing exactly where zero-shot LLM reasoning breaks down in a long-horizon, sparse-reward environment. Balatro is a better testbed than the roguelikes used in prior work because every decision is discrete and interpretable — you can see where the reasoning fails, not just that it does.
PythonBalatrobot APILangChainAnthropic APIPandas
Target Venues
AAAI 2027 and IEEE CoG 2027
What Gets Built
↘Text serializer converting full game state to structured prompt (jokers, hand, shop, money, blind target)
↘LLM agent mapping API responses to valid game actions at each decision point
↘50-100 fixed seeded runs per agent for reproducibility and statistical comparison
↘Decision-level analysis: joker selection quality, hand play vs available options, build consistency
Listening
Modern Wisdom · Hormozi
Modern Wisdom · Naval Ravikant
Tim Ferriss · Naval Ravikant
How I got here
Growing up in Trinidad I wanted to be a doctor. Software engineering was not even on my radar. We were not really exposed to it until the last two years of high school.
It started at 14, making scripts to bot games. The moment I realized you could automate things at scale, something clicked. I figured I could positively impact more lives through software than I ever could through medicine, and I would be at the cutting edge of where the world was actually heading.
I became obsessive about tech blogs and startup news for the rest of high school, trying to absorb everything.
It started at 14, making scripts to bot games. The moment I realized you could automate things at scale, something clicked. I figured I could positively impact more lives through software than I ever could through medicine, and I would be at the cutting edge of where the world was actually heading.
I became obsessive about tech blogs and startup news for the rest of high school, trying to absorb everything.
In 2018 my dad gave me a book, Digital Singularity by Kevin Parikh. It made the case that AI and ML would be the defining force of how the world changes. That book put a name to what I had been sensing and locked in my direction.
For my CSEC and CAPE presentations I focused on ML use cases for real-world problems, reading research papers and trying to explain the ideas clearly to my year group. That drive pushed me to UofT.
Now I am still chasing the same thread. Agentic AI, building things in spaces I genuinely care about, and trying to stay close to where the field is going rather than where it has already been.
For my CSEC and CAPE presentations I focused on ML use cases for real-world problems, reading research papers and trying to explain the ideas clearly to my year group. That drive pushed me to UofT.
Now I am still chasing the same thread. Agentic AI, building things in spaces I genuinely care about, and trying to stay close to where the field is going rather than where it has already been.
Contact
Get in touch
Open to SWE roles in Toronto — AI, full-stack, or anything genuinely interesting. Happy to talk about a role, a project, or just connect.
Full-time and contract · Toronto preferred, remote considered