Skip to content
— 中文 中文

AI-Native Systems

The systems delivery arm of WillChan.

Systems that run every day.

The bottleneck is never the model. It is the three people copying data between tools at 1am, the dashboard nobody trusts, and the process that only one person knows. Every system here exists because one of those was costing someone real money.

Send a briefFrom $200 — a scoped de-risking sprint before any commitment

Fit

Who this is for, and who it is not.

  • You have already tried AI and it did not stick. There are licences nobody opens and a proof of concept that never reached production.
  • One process costs a specific person several hours every day, and you can name that person.
  • You hold the budget and the authority to spend it. Not "we should look into AI at some point".
  • You want the system handed over, not rented. You intend to own and maintain what gets built.

What this is not

  • You want a demo to show a board. A demo is not a system, and this is not a demo shop.
  • The scope has to stay open. An unbounded first milestone is how a two-week build becomes a two-quarter build.
  • You need a body on a timesheet rather than a defined outcome.
  • You would rather we did the work before you decide whether to pay for it.
  • Hourly body-shopping. I quote per system because an hourly rate makes the interesting decisions financially irrational for both of us.
  • Staff augmentation inside a process I cannot influence. If the architecture is already fixed and I am only permitted to implement it, a contract developer is the cheaper and more honest hire.
  • Screens without a system. If the deliverable is a set of mockups that nobody has agreed how to build, the project will fail at handoff regardless of how good the mockups are.
  • Projects where the honest answer is "you do not need AI yet". I will say so on the first call instead of selling a model into a process that needs a data model first.

Selected work

Systems built to run without us in the room.

Ordered by how fast each one makes a founder sit up, not by how impressive the technology is. Open any card and the arithmetic behind every figure is spelled out, so you can check it rather than take it on trust.

Every system here is built in-house and runs in our own stack. They are not client engagements and no external organisation operates them. That is said plainly because the first question a technical buyer asks is who this was built for, and the answer has to be true.

Capabilities

Three pillars of delivery.

  • SLOT A

    Autonomous AI Workflows

    Integrating LLMs into your existing business logic. From multi-agent orchestration to custom MCP server deployment.

    Deliverables

    • Architecture brief: where the model belongs, where it must not be trusted
    • Custom MCP servers exposing your internal tools to any compliant client
    • Multi-agent pipelines with typed tool schemas and deterministic fallbacks
    • Evaluation harness with a regression set before anything reaches production
    • Cost model per workflow and per seat, with hard ceilings and kill switches
    • Observability: traces, token accounting and failure replay
  • SLOT B

    Design Engineering

    Bridging the gap between Figma and production. Delivering pixel-perfect Next.js/React front-ends with robust data architectures.

    Deliverables

    • Production Next.js App Router implementation, statically rendered where possible
    • Design tokens as typed code, with a single source of truth for both themes
    • Data layer: schema design, caching strategy, and loading/empty/error states that were designed rather than discovered
    • Core Web Vitals budget enforced in CI, not aspirational
    • Accessibility to WCAG 2.2 AA on the flows that matter
    • A component library your team can extend without asking me
  • SLOT C

    UI/UX Systems Architecture

    Moving beyond isolated screens. Crafting scalable design systems and Liquid Glass aesthetic interfaces for complex B2B SaaS products.

    Deliverables

    • Information architecture rebuilt around the operator's actual decisions
    • Density strategy: what is always visible, what is one interaction away, what never ships
    • Interaction model for asynchronous AI work — progress, partial results, interruption, undo
    • Role-based surface design for multi-tenant products
    • Governance: naming, versioning and deprecation rules for the system itself
    • Written decisions log, so the reasoning survives staff turnover

Service directions

What we can build next.

These two are not work we have delivered. They are systems we are equipped to build, and they are labelled directions rather than projects on purpose: a direction becomes a work when it has been run for real and left measurable numbers behind, and not before.

  • Service direction · not delivered

    Omni-Channel Content Derivation Pipeline

    Write once. The pipeline derives the X, LinkedIn, newsletter and social versions automatically, then queues them for one human confirmation.

    What this replaces

    A content team’s time disappears into rewriting the same piece four ways, then logging into four platforms to publish it. The breakpoint sits between "finished writing" and "published" — which is the part nobody planned for and the part that scales worst.

    Architecture

    Trigger: a new long-form piece lands via RSS, git, or a manual drop. Brain: an LLM distils platform-specific versions — length, register, hashtags and call to action each set per platform — and then runs a fact-consistency pass so a derived draft cannot drift from the source. Action: a webhook pushes into a publishing queue, and nothing goes out until a human confirms it.

    Guardrails

    • Fact-consistency lock, by design: a derived draft may only extract facts from the source and is forbidden from introducing a new number. This is the discipline already applied to this site’s own published figures, so it is a rule we operate under rather than one we would impose.
    • Hard platform length validation. Over-limit drafts are rejected, never silently truncated — a truncated post is worse than a missing one.
    • Queued is not published. A human gate sits between the queue and the platform, with no auto-publish path.

    What you would receive

    • Workflow blueprint: the n8n / Dify JSON, importable into your own cloud.
    • Platform template library: one template per channel, versioned.
    • Publishing queue dashboard and its interface documentation.

    Stack: n8n · Dify · LiteLLM · PostgreSQL / Supabase · Next.js

    Who this is for D2C brands going overseas, SaaS content teams, and independent developers — anyone producing once and needing to distribute at zero marginal cost.

  • Service direction · not delivered

    Inbound Lead Triage & Auto-Scoping Radar

    A brief arrives. The system produces a company profile, a fit score and a first-draft scope of work before you have opened the email — so your reply in ten minutes reads like a reply you spent a day on.

    What this replaces

    At a high contract value, the quality of the first reply depends on how free the person answering happened to be that afternoon. The breakpoint sits between "brief received" and "substantive reply sent", and it is usually one to two days long — by which point the buyer has already talked to someone else.

    Architecture

    Trigger: the brief form on this site, or an inbound email or message. Brain: fetch the company website and public information to build a profile, score it against the ICP rules, then generate a first-draft execution frame including scope, assumptions and risks. Action: push an interactive card to Slack or Feishu; a human presses [Approve] and the reply goes out.

    Guardrails

    • The AI drafts; a human sends. Same red line as the outbound pipeline — no automated message goes to a prospect, ever.
    • A draft scope must mark its own assumptions. Anything unconfirmed is written as unconfirmed rather than stated as a conclusion, so an unverified guess can never harden into a promise.
    • Public information only. Company profile data comes from public sources; private or personal data is out of scope by design.

    What you would receive

    • Enrichment pipeline blueprint.
    • ICP scoring rules table, editable without touching code.
    • Scope template engine and the approval card flow.
    • CRM write-back mapping.

    Stack: n8n · company data APIs · LLM extraction · Dify / Qdrant · Slack / Feishu cards

    Who this is for Consultancies, agencies, high-ticket SaaS and overseas service providers — where response speed converts directly into close rate.

How an engagement starts

How an engagement starts

Fixed-scope phases with a written architecture brief before any code. Pricing is quoted per system, not per hour, so scope creep is a shared problem rather than a hidden invoice.

  1. 01

    Architecture brief

    A short written document before any code: the bottleneck, the decision that unlocks it, the shape of the system, and the order-of-magnitude cost. It is yours even if we stop here.

  2. 02

    Fixed-scope build phase

    Two to ten weeks, one written scope, weekly working software. No phase ends without something running in a real environment against real data.

  3. 03

    Handover and operating loop

    A decisions log, the runbooks, and a recorded walkthrough. Then a retainer only if the system genuinely needs ongoing tuning — not by default.

FAQ

Questions worth answering properly.

How is this different from a normal AI agency or outsourcing shop?
You receive a system you can hand over: the workflow definitions, the agent prompts and output schemas, the runbook, and the decision log explaining why each part is shaped the way it is. Not a demo, and not a repository only the author can read. The second difference is harder to fake — we build tools for ourselves first, and the proof is on this page. An agency that has never operated a system it depends on is guessing about yours.
How long does a system take to go live?
Two weeks per system, one vertical at a time, is the planning figure. That is an estimate rather than a measurement — the fastest honest answer is that scope is fixed before the clock starts, because an unbounded first milestone is how two-week projects become two-quarter projects. The $200 scoped sprint exists to fix that scope before you commit to anything larger.
Our processes are messy. Can they even be automated?
Sometimes not, and the $200 sprint is designed to find that out cheaply. It produces a fixed-scope architecture brief that names the failure paths first — the integration that will not hold, the data that turns out not to exist, the step that is manual for a reason nobody wrote down. Automating a process nobody can describe produces a faster version of the wrong process. If the honest answer is that AI is not the right tool here, that is what the brief says.
What stops the AI from making things up?
Concretely, and this part is engineering rather than intention: structured outputs are constrained by JSON Schema so a malformed answer cannot be parsed into an action; every generated item carries a confidence score and anything below threshold routes to a human review queue instead of proceeding; retrieval-backed agents answer only from a supplied knowledge base, so a critique cites a rule rather than an opinion; and where a step matters, the pipeline stops on a human [Approve] rather than continuing on its own judgement.
Do you still support it after delivery?
Yes, through an AI Systems Retainer, and that is a position rather than an upsell. These systems decay: a model gets deprecated, an API changes its rate limits, a source page changes its markup, and a workflow that ran untouched for six months stops one morning. A system with no maintenance agreement is a system with a known expiry date. That is also why the name is Systems and not Service — the maintenance is the point.
What have you actually built?
Systems we built and run ourselves, each published with its architecture and its figures. The strongest is an outbound pipeline that runs 1,920 leads a day against a manual baseline of 32 — the throughput of about 60 sales development reps. The weakest, by design, is a Socratic tutoring engine with no performance numbers at all, because the honest claim it makes is about how it teaches rather than how fast. That range is the point: we publish what we can prove and say so when we cannot.
Why is so much of this built in-house instead of shown as client work?
Because it is true, and because it is a stronger claim than a logo wall. These are systems we built and run for ourselves; no external organisation operates them, and this page says so on every card. That is verifiable in a way a client logo is not — anyone can assert a Fortune 500 engagement, but a pipeline that processes its operator’s own leads has to actually work or the operator notices. Client work is published only when the contract allows it, and there is no arrangement under which we would imply one that does not exist.

Secured & escrowed via Upwork Direct Contracts.

Working languages: EN / 中文

Send the brief. Get a written answer.

Describe the problem in a form. What comes back is a specific written response about your situation, including a straight answer when the right move is not to build anything.

Send a brief

If you would rather have a conversation than write, the scheduler is here. Book a Call

Ready to build a system?[ Send a Brief ]