The Back Room
How the system actually works
This is the part no other portfolio has — the operating model behind eight shipped products.
The Framework
Toki Framework v3.4 — ASSESS → ARCHITECT → PLATFORMIZE → ACCELERATE
A four-phase AI product methodology. Every principle was earned from a real production failure, not planned.
Problem definition, idea validation, framework genesis
Milestone
🏁 Toki Framework v1.0 born from SugarStable (Feb 10, 2026)
Built in this phase
8
Sessions
42
Hours
680K
Tokens
Real talk — what hurt
- ↳ Wrong platform choices
- ↳ Scope creep
- ↳ 3 restarts
The Operating Model
Multi-agent lanes with defined scopes
Agents don't roam free. Each runs in a bounded lane with a single writer and a verifier watching.
Architect
Defines the system and writes the spec
Reads the problem, proposes structure, writes the blueprint. Does not write production code.
Writes
Builder
Single writer of production code
The only agent allowed to write to the codebase. Takes specs from the Architect, ships increments.
Writes
Verifier
Reads everything, writes nothing
Audits the Builder's output against the spec and the principles. Can block a ship; cannot edit code.
Writes
Accelerator
Runs the bounded loop
Dispatches the next increment, enforces the loop budget, and escalates to the human when the budget is spent.
Writes
Working Principles
The rules that keep the agents honest
Three principles, enforced in code, that separate demos from production.
Verify, don't trust
Every agent output is checked by a reader-agent before it's allowed to persist. No agent writes and validates its own work.
The Verifier lane reads the Builder's diff against the spec and the principles. If it doesn't match, the write is blocked — not auto-fixed. Trust is earned by passing the check, not assumed by default.
Single writer
Exactly one agent is allowed to write to production code at a time. Everyone else reads.
Concurrent writes from multiple agents produce silent merge conflicts and hallucinated state. The Builder lane holds the pen; the Architect proposes, the Verifier audits, the Accelerator schedules. One pen, many eyes.
Bounded loops
Every agent run has a fixed budget of iterations. When it's spent, the loop escalates to a human — it never spins forever.
A loop counter caps each task. If the Builder can't converge inside the budget, the Accelerator stops the run and surfaces the state to a human. No infinite retries, no quiet token burn.
Build in public
Bugs found & fixed, lessons logged
An honest changelog. The mess is part of the method.
Bugs caught before users did
serverUrl bug at 2am
PLATFORMIZEK-laude.com
What: The chat widget pointed at the wrong serverUrl after a deploy and silently failed.
Fix: Added a runtime origin check and a fail-safe fallback; became the Context-Forward principle.
Base44 auth failure 48h before MLQ launch
ACCELERATEMyLunchQuest
What: Auth broke two days before launch with no code change on our side.
Fix: Pivoted to Lovable + Supabase in 48h; became the Two-Platform Doctrine and Auth Architecture Gate.
LLM hallucination in resume output
ARCHITECTApplyReady
What: Early resume generations invented credentials and metrics that weren't in the source.
Fix: Temperature 0, post-generation validation, anti-fabrication constraints. Became hallucination-resistant design.
4 marketplace ToS violations caught pre-submission
PLATFORMIZEAI Agent Library / Templates
What: Near-misses that would have cost the marketplace accounts if shipped as-is.
Fix: Built a compliance pre-flight check into the framework; zero account issues since.
VoyageHQ interests missing post-rebuild
ACCELERATEVoyageHQ
What: A rebuild dropped the user's saved interests from the year planner.
Fix: Added state-migration checks; became the Verify-Don't-Trust principle for agent writes.
DNS propagation across 3 iterations
PLATFORMIZEK-laude.com
What: Three deploy cycles lost to DNS caching and record misconfiguration.
Fix: Standardized a deploy checklist; baked into the PLATFORMIZE phase.
Lessons that became framework versions
Validate the idea before building. FIT Method: Feature, Integrity, Trust — context and intent defined upfront.
Auth Architecture Gate: validate auth before building any features. temperature 0 for deterministic outputs.
Two-Platform Doctrine: Base44 = Workshop (prototypes), Lovable + Supabase = Factory (production).
Marketplace Intelligence — compliance baked in pre-submission, not after rejection.
AI Leads, Human Decides. Context Forward: the system pre-fills everything it already knows. Simplest Flow Wins.