>_
Hi, I’m Sylvester Wong

I turn AI into a software engineering superpower.

I build production AI systems, developer tools, and practical methods that help teams use coding agents with better context, memory, workflows, testing, and review.

>_ MY CLAUDE CODE GUIDE>_ EXPLORE MY WORK
Production AI Systems
Built and shipped in real environments
Open Source Builder
Recall (active) · Crispy (archived)
Engineering Impact
LLM accuracy 70 → 95% · review time -50%
Based in Toronto
Available for Senior / Lead opportunities
Sylvester Wong
Ask me anything about how I build with AI.
● MODEL OFFLINE
❯
(↵ to submit)
CURRENT WORK

Systems that shipped. Methods that stuck.

$ git log --author="Sylvester Wong"

Recall

Local memory for Claude Code and Codex. Every session saved verbatim and indexed — SQLite FTS5 keyword search fused with local semantic embeddings, so agents can search, read, and continue any past conversation. No cloud, no LLM calls.

TypeScriptSQLite · FTS5llama.cpp
recall.thesylvester.ca →

Crispy

Multi-vendor agent harness orchestrating Claude Code and Codex through one adapter contract, across five surfaces — VS Code, browser, CLI, desktop, and a Discord bot with remote agent control. Resume, fork, and cross-vendor fork any session mid-conversation.

Agent SDKCodex CLIVS CodeTauri
crispy.thesylvester.ca →
◈ fusionthinkin daily use

Fusionthink

Multi-vendor adversarial review for coding agents. A Claude reviewer and a Codex reviewer take the same brief, every claim is verified against the real code, weak findings get pushed back, and disagreements are settled across vendors. Distilled from a review loop run 186+ times in my own sessions.

Claude CodeCodex CLIAgent Skills
thesylvester.ca/fusionthink →
MY CLAUDE CODE GUIDE

How I actually use AI to build software.

The working method behind everything above — context-isolated task handoffs, self-testing sub-agents, automated review validation. Adopted as a 5-person team’s default workflow; review resolution time cut by 50%. Written up as notes, tips, and tools you can use.

research→plan→delegate→review→retain
>_ MY CLAUDE CODE GUIDE
Antidote Health · 2026–present
Australian telehealth platform — patient booking, Stripe payments, EHR/FHIR integration
Claro Customs AI · 2024–25
Conversational AI for a Canada/US customs platform — document ingestion accuracy 70% → 95%, e2e test cycles 15 → 5 min
PromptCore · 2024
Human-in-the-loop LLM evaluation; workflow runner >50% faster than Microsoft Prompt Flow
LighthouseAI · 2023–24
Agentic workflows lifted GPT-3.5 to GPT-4-level classification (70% → 90%+)
VS Code · open source
UX enhancement PR merged into the mainline release
thesylvester.ca⎇ mainAI SYSTEMS · DEVELOPER TOOLS · AGENTIC WORKFLOWScontext loadedTOP--:--:--