Last updated: 25 July 2026
AI Systems Engineer | Creator of Project Athena | Singapore 🇸🇬
I build agentic AI infrastructure that turns generic LLMs into personalised operating systems — own the state, rent the intelligence.
Website: winstonkoh87.com
| Project | What It Does | Engineering Highlights |
|---|---|---|
| Athena | Open-source cognitive augmentation layer — persistent memory, structured reasoning, full data ownership across any AI model (Gemini, Claude, GPT) |
• Chunk-level hybrid RAG (BM25 + semantic + RRF fusion + cross-encoder rerank) • Hit@5 0.908 / MRR@5 0.796 on a 65-query gold set, @ $0 infra (Supabase free tier) • 414 reusable protocols, 72 slash commands, 43 skills • Conditional skill activation (~40-60% token savings) • 1M+ Reddit views + 960+ unique cloners at launch • 7 IDE integrations, MIT licensed |
| Portfolio | 11-page performance-first portfolio with 26 articles and 6 live demos |
• Astro 5.0 + Tailwind CSS (zero-JS islands) • Zero-JS-first architecture, 99/100 Lighthouse |
Languages & Frameworks
AI & Data
Tools & Infrastructure
A platform-agnostic cognitive augmentation layer. Own the state. Rent the intelligence.
| Metric | Value | What It Means |
|---|---|---|
| Sessions | 2,090+ | Continuous context across 200+ days of bilateral use |
| Protocols | 414 active | Open-sourced decision frameworks (reasoning, risk, execution, research) |
| Hybrid RAG | Hit@5 0.908 | Chunk-level BM25 + semantic + RRF fusion + cross-encoder rerank — MRR@5 0.796, coverage 0.639, measured on a 65-query gold set |
| Skills | 43 clustered | Cognitive Cluster architecture — co-activated skill pipelines with conditional activation |
| Case Studies | 489 | Documented friction → solution patterns with empirical outcomes (post-dedup) |
| Scripts | 259 | Python automation stack (boot, shutdown, search, sync, hooks) |
| Launch Reach | 1M+ views | Reddit launch post — 960+ unique cloners, 4.5K+ unique visitors |
Key Engineering:
- Hybrid Search: pgvector + BM25 keyword + RRF fusion + cross-encoder rerank — Hit@5 0.908 / MRR@5 0.796 on a published 65-query gold set, hardened through logged production regressions
- Conditional Skill Activation: Path/topic-triggered dormant skills reduce prompt bloat by ~40-60% (Protocol 530)
- Epistemic Honesty as a Feature: public Validation Status ladder grades every claim by evidence level; 18 Crossref-verified academic references (APA list); mechanisms labeled code-enforced vs agent-discretion
- Guards That Can Actually Fail (v9.9.8): an audit found five green checks that could not detect the defect they named — a CI step running the writer instead of the check, a version check scoped to the three files that already agreed,
ruff --exit-zero, tests asserting a return type instead of a behaviour, and a privacy scanner excluding its own config. Standing rule now: any fix to a guard must show it failing before it shows it passing, mutation-tested and recorded in the commit - Multi-Agent Orchestration: Coordinator synthesis discipline with anti-delegation enforcement and token budgeting
- Atomic Writes: POSIX-compliant data safety for all memory operations
- Privacy Pipeline: CI privacy gate over 41 patterns, with the committed blocklist restricted to shape-only matchers — a public blocklist that names people is an index of what you're hiding, so name patterns live in a repo secret and the scanner audits its own config
- Semantic Cache: LRU with disk persistence, cosine matching, and Supabase delta sync
- Zero Infrastructure Cost: Runs on Supabase free tier + local compute. No cloud bills.
| Capability | Evidence |
|---|---|
| RAG Pipeline Engineering | Production hybrid search: BM25 + semantic + RRF fusion + rerank. Hit@5 0.908, $0/month infra |
| Agentic AI Systems | 414 protocols, 72 workflows, 43 skills — full agent lifecycle (boot → work → shutdown) |
| Multi-Agent Coordination | Parallel worktree orchestration, coordinator synthesis, conditional skill activation |
| Full-Stack Web Development | 7 production sites, Astro, zero-JS-first architecture |
| AI Consulting | Active client engagements — diagnostics, AI integration strategy, workflow automation |
| Technical Writing | 26 articles, 9.8K+ views — clear communication of complex systems |
- 9.8K Views, 750 Cloners: The Day I Shipped My Brain to the World
- Why I Built My Own Brain (The 5 Pillars of Sovereign AI)
- The Trilateral Feedback Loop: Why One AI is Not Enough
- The Bionic Operator: Why AI Replaces Tasks, Not Humans
- The Anti-Slop Protocol: How to Write 3,000 Words in 3 Hours
- The Vibe Coder's Trap: Why AI Speed Can't Fix Business Physics
Looking for pragmatic builders who:
- Ship fast, iterate, break things at 70% readiness
- Value robustness over cleverness
- Are comfortable with async communication
📬 Reach out: [email protected] or DM on LinkedIn
"The best way to predict the future is to build it."