Engineering Record · Waryl Systems
We run our business on the systems we build.
Why not build this in-house? We asked the same question, then built these systems for ourselves. This page is the engineering record of what operates Waryl today: research pipelines, an editorial engine, trend detection, security audits. Read the decisions first. They are the part a build budget does not include.
The architecture: what the diagram shows
Read the stack top to bottom: a company, not a script.
This diagram shows the system that operates Waryl today.
One company. Seven layers. Final say: human.
- 01
Sensing
intelligence · LLMLLM5 sources · weekly collector
Trends MCPYouTube APIYouTube SuggestGitHubHacker News5-10 API keys · rotation - 02
Strategy
planning · LLMLLMADR-12 · priority rules
Orchestratorpriority rules5 engines: Trend · Evergreen · Strategic · Personal · Derivativeweekly cycle · 6 phases · feeds next cycle
- 03
Execution
operations · LLM + deterministicLLM + DETbuild ops · 10-12 agents
10-12 agents per operationcopy · animations · nav · trends · research - 04
Governance
quality · deterministic + LLMdeterministicreview log · accept/modify/reject
Gate · Devil's Advocatescoring below 4.0 rejectsacceptmodifyrejectForge 2.1/5 · rejectedHeartbeat 4.3/5 · shippedcopy 11/7/112+ ADRs5 agents + DA · 5 rounds - 05
Memory
institutional · deterministicdeterministic~107 docs · updated this week
internal wiki · ~107 docsindex.jsonupdated this weekcode is a derivative of the docs — the index is the truth
- 06
Security
defense · deterministicdeterministic17 to 0 · audit Jul 2026
17 → 0 vulnerabilities8 leaf agents · in parallel3 protection layers - 07
Leadership
authority · humanhuman reviewOscar, the architect
decision note · ADR-012
approves in person · the system suggests, the architect decides
Not a session. A company.The same pattern closed 17 vulnerabilities to 0, with 8 leaf agents auditing in parallel.
- 01 · Sensing. Five sources feed the intelligence layer: Trends MCP, YouTube API, YouTube Suggest, GitHub, Hacker News. API keys rotate across free-tier accounts so nothing rate-limits.
- 02 · Strategy. Five engines plan the weekly cycle: Trend, Evergreen, Strategic, Personal, Derivative. An orchestrator applies priority rules and decides what runs when. Agents never self-select.
- 03 · Execution. Build operations run ten to twelve agents per operation: copy, animations, navigation, trends, research.
- 04 · Governance. A gate scores every output. Below 4.0 rejects: The Forge at 2.1/5, Heartbeat Monitor at 4.3/5, copy validation 11 accepted, 7 modified, 1 rejected. No output skips this gate.
- 05 · Memory. Every run writes to one source of truth: index.json for the editorial engine, an internal wiki of ~107 linked documents. The index is the truth; code is a derivative of the docs.
- 06 · Security. Eight leaf agents audited the stack in parallel and closed 17 vulnerabilities to 0, across three protection layers.
- 07 · Leadership. The final delivery is approved by a human, in person. Documented principle: the premium user decides, the system suggests.
Data packets travel the rail; the gate is the node no packet passes without a verdict. The same pattern closed 17 vulnerabilities to 0, with 8 leaf agents auditing in parallel.
Three decisions, with trade-offs
1. Orchestrator with priority rules, not free choreography
Decision
The editorial engine runs 5 recommenders (Trend, Evergreen, Strategic, Personal, Derivative). An orchestrator applies priority rules across a weekly 6-phase cycle. Agents never self-select.
Trade-off
Choreography is flexible; rules are predictable. We gave up "agents figure it out" for "we can state what runs next Tuesday." Adding a source means editing rules, not adding an agent.
What we learned
With free choreography you can only describe what happened; with rules you can specify what will happen. For weekly operations, predictability is worth the rigidity.
2. Deterministic first; LLM only where judgment is required
Decision
The site is static Astro with zero unnecessary JavaScript. The hero demo is a deterministic SVG blueprint, no LLM involved. Supabase only receives leads. LLMs are reserved for synthesis and validation.
Trade-off
Deterministic paths are cheap, fast, verifiable. LLM paths are probabilistic and slow. Removing the LLM from the demo costs spectacle; putting it where a rule suffices costs trust.
What we learned
A demo that behaves identically on every visit builds more credibility than a clever one. The LLM's value is judgment, not rendering.
3. Failure is an input, not an accident
Decision
Hypotheses scoring below 4.0 are rejected. The Forge animation was rejected at 2.1/5; Heartbeat Monitor shipped at 4.3/5. Copy validation: 11 accepted, 7 modified, 1 rejected. The trend pipeline rotates 5 to 10 free-tier API keys across 5 sources (Trends MCP, YouTube API, YouTube Suggest, GitHub, Hacker News). Every deliverable passes an accept/modify/reject gate. Postmortems are kept, including "7 fatal failures of the Devil's Advocate."
Trade-off
Gates and scoring add latency and cost; rotation adds operational overhead. Without them, a silent pass ships confidently wrong and one rate limit kills the week.
What we learned
A rejection with a score is more useful than an approval without one. Design for the failure mode: record it, score it, feed it back. The same audit removed 473 AI-typical em-dashes from our visible copy; we run that check on everything, including this page.
Stack
Astro 7, Tailwind CSS, GSAP, Lenis, Supabase (leads only). Agent operations run on LLM providers: OpenAI, Anthropic, and open-weight models. Plain names, no badges: the architecture is the product, not the libraries.
Bring your architecture. We will contrast it against this one, point by point, and tell you where in-house makes sense. Sometimes it does; we will say so.
Free architecture review · 15 minutes
No pitch. Point-by-point contrast. Sometimes the answer is in-house; we will say so.