Building systems that reason
REASON
DSPy optimization, RAG pipelines, skill-based agent workflows. Making LLMs do useful work reliably.
- —Prompt optimization & evaluation
- —Content generation pipelines
- —Skills as structured pipelines for subagents
REMEMBER
Postgres-native architectures. pgvector, JSONB, semantic search without separate vector or graph DBs.
- —Vector embeddings in Postgres
- —Event-sourced state & temporal queries
- —JSONB for flexible metadata
What I'm Working On
Research Intelligence Platform
ACTIVEIngests hundreds of articles and research papers daily, ranks by relevance and novelty, and ensures editorial variety before publishing. Includes a RAG-powered research assistant querying the full news and research paper knowledge base.
DSPy pipelines, isolated subagents with reusable skills, Postgres with pgvector for semantic search. Powers theqi.news and thelongview.news.
Six Lines
COMING SOONNative iOS I-Ching and Chinese almanac built on primary sources—3,600+ scanned pages from Qing imperial manuscripts, character-by-character semantic translations, and 4,096 Yilin transformation verses. Every feature traces back to classical volume-and-page citations.
Multiple scholarly lenses on the same hexagram: decoded commentary, Wilhelm translations, Hatcher semantic matrices. On-device AI via Apple Foundation Models. EN, 繁體, 简体.
8-Bit Oracle
ACTIVEI-Ching divination with a modular architecture—core interpretation and voice are separate layers. Exposing the bare interpretation lets users judge what the AI is actually telling them before any narrator voice dresses it up.
Layered pipeline: casting, interpretation, voice. Three styles (義理 moral philosophy, strategic, practical). Four languages.
Pix
Hackathon WinnerAn autonomous oracle agent on Twitter. Performs I-Ching readings and stores insights on OriginTrail's decentralized knowledge graph. Won Consensus HK 2025.
Augustin Chan
Before AI, I spent 10 years at Informatica building and architecting master data management systems for enterprises across APAC and Europe. Large-scale data architecture, complex integrations, the kind of work where you learn that systems need to be robust before they can be clever.
Now I apply that discipline to AI: prompt optimization with DSPy, domain expertise encoded as reusable skills, memory systems that let LLMs maintain context across sessions. For multi-agent work, I apply concurrency patterns from distributed systems — isolated subagents with skills, no shared state, merge after completion — instead of the “agents coordinating with each other” theater.
BS Cognitive Science (Computation) UC San Diego · Bronx HS of Science
Elsewhere
Talks, interviews, and places I've shown up