13 · Founder, WebCanBe
Archive · 2025–2026
This is the fuller record. I keep experiments that ended, systems that changed shape, and research directions that failed alongside the work that continued.
Started as a long-term attempt to build an assistant that could move beyond chat. The implementation grew into a multi-stage system: speech input → transcription → model/router → memory/context → queued tool actions → confirmation gates → speech or interface output. Later iterations explored local/cloud model routing and computer-facing tools rather than tying the project to one model.
Initial concept for connecting sellers with unused local storage capacity. The idea expanded into a workflow system with seller/host roles, requests, approvals, inventory states, QR movement, and fulfillment handoffs. Implementation work continued into 2026.
Worked from real store data toward a system that could detect operational risk rather than merely display stock. The design included inventory synchronization, safety stock, lead-time-aware stockout logic, reorder proposals, and human approval.
Began a research track around virtual unwrapping and ink detection for carbonized scrolls, with emphasis on geometry integrity, reproducible pipelines, intermediate inspection, and the challenge's evidence constraints.
Started the project that became WebCanBe. It would move through service, marketplace, production, and software-workspace forms rather than staying fixed to the first version.
The project began moving away from the earliest model toward a more explicit web-production and productized-delivery direction.
A focused service experiment around creating and shipping storefronts. Useful mainly because it exposed repetitive handoff, implementation, and revision work that later informed product thinking.
Started a dedicated research log around the Beal Conjecture. The project developed a strict evidence discipline: primary sources, explicit hypotheses, case reductions, modular/Frey routes, multi-Frey literature, height/S-unit bounds, and a state system distinguishing known results from derived or candidate steps.
Another deliberate repositioning changed how the project thought about delivery, customer workflow, and where repeatable value lived.
Built a working SaaS flow for idea analysis, risk surfacing, execution planning, comparison, and progress tracking. The prototype was built fast enough that market distribution became the next experiment instead of more feature work.
Explored a Shopify-oriented return/refund workflow where deterministic rules and AI suggestions were constrained by thresholds and human approval. The useful idea was not “AI refunds”; it was human-in-the-loop decision automation for actions with real cost.
The current product thesis began taking shape: the real codebase should remain the primary object while visual editing, direct code editing, AI changes, preview, export, and deployment become different interfaces to the same source.
Defined a broader end-to-end product flow where projects could be discovered or acquired and then continue living inside an editable workspace instead of ending at download.
Designed explicit rules for classifying framework/runtime boundaries before allowing aggressive visual edits: server/client boundaries, environment-variable exposure, unsupported patterns, and a scanner-oriented compatibility model instead of pretending all repositories are equally safe to transform.
Built an iOS/Swift control surface for sensitive administrative operations, including manifest validation, operation rendering, and XCTest coverage. The larger idea was to separate privileged actions from ordinary product UI and make them inspectable.
Worked through authentication and security edge cases across Firebase-based flows. Recorded validation passes included a 2,087-test password-reset run and a 940-test emulator pass around passkey work, alongside contract and browser checks for route, secret, and preview-injection boundaries.
Built and refined Git/GitHub/Vercel workflows around recoverable projects, source ownership, deployment, and failure recovery. One important principle came from real repository recovery work: a project is not truly owned if its current working state cannot be reconstructed when a tool, sync layer, or environment fails.
Structured prospecting around researched accounts, verified addresses, source URLs, deduplication, one clear CTA, explicit opt-out, send-state tracking, and deliverability-aware infrastructure. The workflow treated outbound as a data-quality system rather than a mail-merge problem.
Tested multiple customer frames to learn where production capacity, source ownership, and faster shipping were painful enough to matter. The point of each segment was not to accumulate personas; it was to eliminate weak positioning faster.
A project can be small in revenue or users and still be technically serious. I try to keep the record focused on what was actually built, learned, or ruled out.