Commercial and operational truth stays outside model prose.
Selected work
What I do, and the proof behind it.
The projects stay the same. Pick the capability you care about, then inspect the decisions and evidence behind the work.
AI-powered apps
The app is the easy part. The system around it is not.
These products hold conversation state, bound what a model may decide, keep commercial truth in a database, and stay operable after launch. Each one reached real users, not a demo.
Candidate sets are bounded before a planner can select.
Safety, eligibility, and human takeover remain explicit.
Claims are tied to a named verification surface.
AI-powered apps
Each record starts with the operating problem and ends with evidence.
Muriersilk AI Sales Concierge
Production deployed / live customer surface
A reusable commerce concierge that streams through CopilotKit and AG-UI, retrieves products and knowledge through a guarded hybrid pipeline, hydrates current commercial facts in the browser, preserves durable context, and hands operations to Chatwoot.
The operating problem
A useful commerce assistant must help with discovery and care questions without inventing prices, recommending ineligible products, turning marketing copy into medical claims, or coupling model failures to checkout.
What I owned
I defined the product boundary, retrieval architecture, two-run interaction contract, context budgets, catalog and safety guards, multi-tenant runtime, operator workflow, and release evidence. AI agents produced most implementation code under those constraints.
The decisions that mattered
- Keep PostgreSQL and Medusa as the commercial source of truth while treating Qdrant as a rebuildable retrieval projection.
- Let the model select stable product IDs and reasons, then let the browser hydrate current names, images, prices, currency, and URLs.
- Compile planner and final-generation context below a 28k hard budget so provider input remains below the 32k pricing boundary.
What can be checked
- A real provider smoke completed Run A, exact glove cards, Run B, a second product-card turn, and replay. The highest observed compiler input was 25,822 tokens and the highest provider prompt was 5,785.
- Backend fresh verification passed 70 test suites and 572 tests, followed by build:chatbot.
OptiTalent
Independent product / active validation
A multi-tenant recruitment product that turns candidate search, evidence review, shortlist decisions, controlled sharing, and recruiter feedback into one coherent workflow.
The operating problem
Recruiting teams need to move quickly without turning matching into an opaque score or losing the evidence behind a decision.
What I owned
I defined the product boundaries, tenant model, evidence contracts, recruiter workflow, review gates, and release acceptance. AI agents produced most implementation code under those constraints.
The decisions that mattered
- Prefer evidence-first retrieval and explicit unknowns over a single opaque LLM judgment.
- Place bounded embedding and retrieval work close to tenant data while preserving server-side policy and review authority.
What can be checked
- Implemented candidate search, candidate evidence, shortlist, and sharing contracts.
- LightGBM LambdaRank was evaluated against deterministic baselines on a frozen retrieval protocol and was not selected as the production winner for that feature shape.
Piveo
Public open-source repository
An open-source creative workstation for image and video generation workflows, designed around repeatable creative settings, queues, galleries, comparison, and human control.
The operating problem
Creative generation becomes difficult to operate when prompts, media, retries, settings, outputs, and comparisons are scattered across one-off chats.
What I owned
I defined the workstation product, generation flow, queue behavior, creative controls, gallery, cancellation and retry states, and self-hostable release shape.
The decisions that mattered
- Treat generation as a controllable workflow with state, not a single prompt box.
- Keep outputs and settings inspectable so people can iterate deliberately.
What can be checked
- Public repository contains image and video generation surfaces.
- Runtime includes a generation queue, gallery, download, comparison, retry, and cancel behavior.
AI automation
Automation keeps its human signature.
I map who approves what, which steps can run unattended, and where evidence must attach to a decision. The workflow gets faster without losing accountability.
Start with the operating problem, the guardrails, and the decision that still needs a person.
Domain rules sit below every interface, whether the entry point is CLI, MCP, or a browser.
The interface makes drafts, approvals, exceptions, and recovery visible.
Every automated output carries the evidence a reviewer needs to accept or reject it.
Automation records
Document-heavy workflows automated with review gates a person controls.
Tender Luban
Private client delivery
A private tender workflow that helps a company assess whether its qualifications, personnel, certificates, and verified past performance are sufficient for a bid.
The operating problem
Tender teams spend significant time checking requirements against scattered company evidence before they can decide whether to bid and what can safely be assembled.
What I owned
I translated operating knowledge from procurement and tender work into capability boundaries, human review gates, evidence selection, audit state, and commercial document export.
The decisions that mattered
- Treat company evidence as selected and reviewed inputs, not free-form model memory.
- Keep high-impact tender decisions behind human approval.
What can be checked
- Qualification, personnel, and past-performance sections are implemented in the commercial bid exporter.
- The private product exposes commercial bid selection and export workflows.
Muriersilk AI Sales Concierge
Production deployed / live customer surface
A reusable commerce concierge that streams through CopilotKit and AG-UI, retrieves products and knowledge through a guarded hybrid pipeline, hydrates current commercial facts in the browser, preserves durable context, and hands operations to Chatwoot.
The operating problem
A useful commerce assistant must help with discovery and care questions without inventing prices, recommending ineligible products, turning marketing copy into medical claims, or coupling model failures to checkout.
What I owned
I defined the product boundary, retrieval architecture, two-run interaction contract, context budgets, catalog and safety guards, multi-tenant runtime, operator workflow, and release evidence. AI agents produced most implementation code under those constraints.
The decisions that mattered
- Let the model select stable product IDs and reasons, then let the browser hydrate current names, images, prices, currency, and URLs.
- Share one capability core across dedicated and multi-tenant profiles while routing from trusted origins and Chatwoot accounts with fail-closed behavior.
- Keep business prompt editing in Chatwoot while platform safety policy, tool schema, and tool permissions remain outside operator control.
What can be checked
- A real provider smoke completed Run A, exact glove cards, Run B, a second product-card turn, and replay. The highest observed compiler input was 25,822 tokens and the highest provider prompt was 5,785.
- Storefront verification passed 56 test files with 860 unit tests, four Playwright browser scenarios, and the production build.
Full-stack web delivery
A site is a system, not a page.
These deliveries pair information architecture and technical SEO with commerce integrations and edge deployment. Released, measured, and iterated after launch.
A page owns a specific commercial question.
Routes stay legible before templates scale.
Data drives metadata, schema, locale, and market context.
Growth work cannot destabilise commerce or delivery.
Technical and commercial signals inform the next bounded decision.
Full-stack web cases
Information architecture, commerce, and controlled releases tied to commercial intent.
Muriersilk
Live client delivery
A modern D2C silk-commerce storefront with an evidence-gated operating loop for technical SEO, content quality, indexation, and commerce-safe delivery.
The operating problem
The storefront needed modern frontend and cloud delivery while recovering search visibility without risking checkout, pricing, shipping, or production commerce behavior.
What I owned
I owned the storefront architecture, migration and delivery decisions, SEO evidence gates, release verification, and the boundary between growth work and commerce risk.
The decisions that mattered
- Rank SEO work by commercial value and evidence rather than producing an unbounded content backlog.
- Keep checkout and commerce behavior outside routine SEO changes.
- Require live browser and production evidence before calling a release complete.
What can be checked
- Public storefront at muriersilk.com.
- The product runtime includes sitemap generation, structured-data helpers, and reproducible SEO evidence gates.
Kinmed
Live client delivery
A multilingual B2B medical-device web platform combining structured product information, country and category discovery, programmatic SEO, and Cloudflare-oriented delivery.
The operating problem
An export-focused medical supplier needed buyers in different markets to find relevant products and regulatory context without maintaining each locale and market page by hand.
What I owned
I designed the information architecture, country and category page matrix, metadata and hreflang behavior, structured content model, and release verification workflow.
The decisions that mattered
- Design the site tree before templates so one canonical page owns each intent.
- Generate pages from structured product and market data while preserving indexation quality gates.
- Keep localized structured data aligned with visible page content.
What can be checked
- 1,344 geo-localised pages across country, category, and locale combinations.
- Public website at chinakinmed.com.
Open source & contributions
Public review is stronger evidence than a private claim.
I separate projects I own from contributions another maintainer accepted. Both matter, but they prove different things.
Independent systems and reviewed contributions remain clearly separate.
A maintainer merge is recorded as external confirmation, not self-attested proof.
Each record states the contract, the repository, and the boundary of the change.
Public links make the source, review, and current state inspectable.
Open source
Public systems and framework contributions that survive review.
Public systems
Evidence-oriented runtimes and controllable AI tools.
Open source is where architecture, onboarding, contracts, and public evidence have to remain inspectable without private context.
External contributions
MERGEDCloudflare vinext #2902data-returning Server Action state-preservation contractMERGEDCloudflare vinext #1144App Router typegen across route graph, CLI, and Vite devMERGEDCloudflare vinext #646request-scoped cache lifecycle primitiveMERGEDrust-ui #47lifecycle-safe Leptos shortcut with browser regression coverageMERGEDpi-agent-flow #28smart mode for single-purpose tasksOPENVespa #37339JSON summary output for inspect profile
Claim discipline
Subsystem ownership is earned through review and maintenance.
I label ownership and contribution state precisely, and keep framework work distinct from the central runtime.
Public projects
Projects I own are separate from contributions another maintainer reviewed.
Memtruth
Public open-source repository
An open-source SDK and runtime for evidence, parsing, corpus contracts, chunking, projection, diagnostics, provenance, and citation-aware AI applications.
The operating problem
AI applications often turn documents into untraceable text, making it difficult to inspect where a field came from, what parser produced it, or whether a result can be audited.
What I owned
I shaped and contributed evidence, provenance, citation, parser, CLI, and release contracts across the public repository and its runtime integrations.
The decisions that mattered
- Keep evidence explicit instead of burying source context in prompts.
- Use schema-versioned public contracts for outputs that other systems consume.
What can be checked
- Public repository includes explicit citation and provenance contracts.
- Tests cover evidence-aware extraction and readable audit output.
Piveo
Public open-source repository
An open-source creative workstation for image and video generation workflows, designed around repeatable creative settings, queues, galleries, comparison, and human control.
The operating problem
Creative generation becomes difficult to operate when prompts, media, retries, settings, outputs, and comparisons are scattered across one-off chats.
What I owned
I defined the workstation product, generation flow, queue behavior, creative controls, gallery, cancellation and retry states, and self-hostable release shape.
The decisions that mattered
- Treat generation as a controllable workflow with state, not a single prompt box.
- Keep outputs and settings inspectable so people can iterate deliberately.
- Make self-hosting a first-class delivery path.
What can be checked
- Public repository contains image and video generation surfaces.
- Runtime includes a generation queue, gallery, download, comparison, retry, and cancel behavior.