B.Sc. Computer Science & Statistics, University of Toronto (Mathematics minor, 2025). I build AI systems that run in production — LLM pipelines, RAG architectures, anomaly detection, full-stack SaaS. Recently completed research with Prof. Patrick Hosein at TTLab on zero-shot LLM reasoning in long-horizon decision environments, accepted at IEEE ICTMOD 2026.

From Trinidad and Tobago. Based in Toronto.

• Maritime KYC/AML Platform — building a custom compliance platform for a financial services firm, replacing a legacy World-Check integration. Entity resolution over sanctions/PEP data using pgvector + LLMs, with a 4-tier RBAC system and a config-driven risk-scoring engine
• "Can Generative AI be used to win at Balatro?" — accepted at IEEE ICTMOD 2026. Compared heuristic, zero-shot LLM, and RAG agents on long-horizon decision-making under uncertainty, using Balatro as a controlled testbed
• Applied AI systems — production RAG pipelines, agentic workflows, LLM-powered SaaS

Maritime Financial Group (Software Engineer, Contract — Sep 2025 to present)

• Building a metadata-driven KYC/AML compliance platform replacing a third-party vendor system — sanctions/PEP screening with LLM-based entity resolution over pgvector, a config-driven risk-scoring engine, and a 4-tier RBAC system across 5 branches
• Built an end-to-end firewall and VPN anomaly detection pipeline ingesting 100k+ log events per week — automated risk scoring reduced manual analysis time by 70%
• Designed a hybrid detection system combining a rules engine, Isolation Forest, and autoencoder over 5-minute windows, targeting login bursts and multi-region access anomalies with explainable per-alert scoring
• Shipped a Next.js monitoring dashboard with grouped alerts, drill-down views, and trend analytics — cut average analyst investigation time per incident by 40% and reduced repeated false-positive escalations

"Can Generative AI be used to win at Balatro?" (IEEE ICTMOD 2026, accepted)

• Compared five agents — a floor baseline, a hand-coded heuristic, a zero-shot LLM, and two RAG variants — on 347 controlled game trials of long-horizon strategic decision-making in Balatro
• Found zero-shot LLM reasoning matches a hand-coded heuristic outright, and that LLM reasoning over expert strategy knowledge beats deterministic execution of that same knowledge on the paper's finer-grained metric
• Also trained a MaskablePPO RL agent to ~310M timesteps; traced its failure to a reward-misspecification bug in the training environment's shop economy (the policy learned to hoard cash instead of buying upgrades) rather than a flaw in the RL algorithm itself

Inbox Copilot

• Built a production AI email client where reply drafts are grounded in real Gmail/Outlook history via LangChain + Pinecone RAG, with real-time inbox sync via Aurinko webhooks and multi-account support
• Shipped as full SaaS with Stripe subscription billing, free/pro feature gating, and end-to-end type safety via tRPC

AI Developer Collaboration Platform

• Built a RAG platform that ingests GitHub repositories via the GitHub API, indexes source files and commits into pgvector, and answers natural-language questions about any codebase with exact source-file citations
• Added AssemblyAI meeting transcription indexed alongside code context; shipped credit-based SaaS billing with Stripe, deployed via Docker on Fly.io

Prepwise

• AI mock interview platform where a Vapi voice agent conducts real-time interviews, generates role-specific questions via Google Gemini, and scores answers across 5 structured criteria
• Validated across 100+ sessions with Firebase Auth and Firestore persistence

F1 Lap Time Predictor

• Trained an LSTM on Monaco GP 2023 telemetry via FastF1 to predict next-lap time from a 10-lap rolling window; compared against an XGBoost baseline with SHAP explainability
• Tyre degradation and sector consistency were stronger signals than raw top speed — the data made this project inevitable

NBA Hot Hand Bayesian Analysis

• Tested the Hot Hand Fallacy across 15 NBA players using Bayesian regression in PyMC; compared Multivariate Normal, Horseshoe, and Spike-and-Slab priors — reduced model divergences by 20% via Metropolis-Hastings tuning
• Presented model selection rationale, posterior diagnostics, and sports analytics implications to the UofT Statistics department head

CommerceCore

• Production-style e-commerce REST API in Java with JWT authentication, role-based access control, and full cart/order/product management
• Clean layered architecture: Controller → Service → Repository → DTO

Algorithmic Trading Backtest

• Backtested a Twitter engagement-ratio signal against NASDAQ across 2022 — portfolio outperformed benchmark by up to 25%
• Built full signal generation, portfolio simulation, and performance attribution pipeline from scratch

• IBM RAG and Agentic AI — Professional Certificate, 8 courses (LangChain, LangGraph, CrewAI, AutoGen) · Oct 2025
• Generative AI with LLMs — AWS + DeepLearning.AI · Jun 2025
• Machine Learning Specialization — DeepLearning.AI + Stanford, Andrew Ng · Jan 2024

Software Engineering and ML/AI Engineering roles in Toronto — backend-leaning, full-stack, or anything at the intersection of AI and practical software. Especially interested in roles where I can own systems end-to-end.

vinayakcpa@gmail.com · vinayakmaharaj.dev