17 min read
A Waterloo Student's Honest Recruiting Recap 2025/2026
From zero interviews to 30+ rounds including startups, unicorns, and FAANG. The unfiltered version.
- career
- personal
Hi, I'm Rick.
>
From the model to the metal, I build it all.
a stonecutter hammering 100 times with no crack, then it splits on the 101st...
Open to Summer 2027 co-op
UWaterloo
UWaterlooMy experience
 logo](/_next/image?url=%2F_next%2Fstatic%2Fmedia%2Fguidept.06afb089.png&w=96&q=75){kind=small}
AI Engineer
Joined the Agentic PM team building a LangGraph multi-agent system that automates expert-network project management as a five-agent pipeline with human-in-the-loop approval gates.
September 2026 – Present logo](/_next/image?url=%2F_next%2Fstatic%2Fmedia%2Frocket.ea34fdbb.png&w=96&q=75){kind=small}
Software Engineer
Led the initial implementation of a C# transaction verification service for the X1/Westcor ecosystem, extending a RED-only check to three-state RED/YELLOW/GREEN suppression logic with LMBRU underwriter notifications. Contributed to the X1 Extended Update integration automating abstract retrieval across 28 states, realizing ~$1M in annual savings, and migrated CI/CD to Argo Rollouts to enable canary deployments.
January 2026 – April 2026
AI Engineer
University of Waterloo IST
Built JADA, an AI-powered job ingestion and recommendation pipeline in collaboration with Microsoft Canada, designing a KNN-based classifier using Cosmos DB Vector Search, and developing full-stack portal services (resume search, profile-based recommendations, InStage integration) supporting 42,500+ users per term.
May 2025 - August 2025 logo](/_next/image?url=%2F_next%2Fstatic%2Fmedia%2Finspur.1ce32c51.png&w=96&q=75){kind=small}
AI Engineer
Built the RAG document-ingestion pipeline for a low-code LLM application platform, converting heterogeneous documents (PDF, PPTX, Excel) into normalized Markdown, then chunking, embedding, and indexing into a Milvus vector store. Optimized chunking strategy via an evaluation harness, reducing retrieval errors by 30%, and designed an OCR fallback for scanned documents. Curated and labeled training data to fine-tune a Surya OCR model, improving text and table extraction on image-based formats.
September 2024 - December 2024My projects
MindGo
Made with:
A personal finance app built around the academic term rather than the calendar month, windowing every report on real term boundaries. Import transactions by dropping in a screenshot — a bank list, receipt, Uber trip, or WeChat Pay history — read by an OCR model that runs entirely in the browser, so the image never leaves the device. Tracks six currencies with read-time conversion, alongside savings goals, a live investment watchlist, and AI-generated plans grounded in the user's own spending. Market data falls through a three-source chain (Finnhub → Yahoo Finance → Alpha Vantage), so a missing key degrades rather than breaks. Bilingual EN/ZH. 50+ users in production.
SlopFish
Made with:
A fork of Stockfish that plays a chess variant where capturing is mandatory whenever a capture is legal. The rule has to hold at every node, not just the root, so the constraint reaches into move generation, move ordering, check evasions, and the quiescence search — where stand-pat pruning has to be disabled entirely, since a forced capture breaks the lower-bound assumption alpha-beta relies on. Trained a variant-specific NNUE on 17.2M self-play positions, measured it against Stockfish's default network, and shipped the default. A self-play harness whose arbiter enforces the capture rule itself — rather than trusting the engine — surfaced a rule violation in 2.5% of games, traced to promotions being counted as captures; fixed across six call sites, then verified at zero violations in 400 games.
DearByte
Made with:
A self-hosted personal AI agent that combines Apple Watch health data, calendar events, and persistent memory to deliver contextual morning briefs through Telegram. Integrated an open-source HealthKit bridge through remote MCP on Cloudflare Workers, then added daily snapshots, personal baselines, and data-freshness checks. Built a TypeScript agent loop with validated tool calls, model-usage tracking, and enforced spending limits; connects to MindGo through read-only MCP for spending context. Implemented an x402 purchase flow on Base Sepolia using test USDC, with explicit user approval, seller allowlists, spending caps, and payment receipts enforced in code. Open source · Active development.
AGE Game Engine
Made with:
Architected a C++20 module-based MVC game engine for 2D ASCII games, currently supporting Space Invaders and Snake. Implemented core engine systems including rendering, input handling, game loop management, and collision detection. Designed a flexible Entity-Component System (ECS) for modular game object management, enabling easy extension with new entities and behaviors. Games inherit from a base GameEngine class and implement virtual methods for game-specific logic, streamlining new game development.
AI Sentiment Pulse
Made with:
Led the ML work on a 9-member WAT.ai team building a market sentiment analytics system, fine-tuning DeBERTa-v3 for sentiment and sarcasm detection. Aggregated model outputs with unsupervised statistical methods to produce real-time sentiment signals from social media, and correlated those signals against financial data APIs to surface tradeable trends.
My Education
Bachelor of Mathematics
University of Waterloo
2023 - 2027
- Major in Computational Mathematics
- Major in Statistics
- Minor in Combinatorics and Optimization
- Minor in Computing (Computer Science)
- Activities: Core Member at Waterloo Blockchain Club; ML Lead at WAT.ai (Sentiment Pulse); Basketball Intramurals
Blog
7 min read
Echoes of the City of Angels
Story of spending a month’s salary to bridge the gap between a 14-year-old’s childhood fantasy and the raw, beautiful, and fractured reality of Los Angeles
- travel
- personal
Contact me
Please contact me directly at m69liu@uwaterloo.ca or through this form.