Spend reserved, then reconciled
Estimated provider costs are reserved before each request and reconciled afterward, so proposals stay inside permissions and spending limits.
All work01 / 06
AI product · Live product
Distrove.
An AI growth team for solo founders.
Demo workspace with an example business. Switch to System to see the architecture.
StageLive product, in active development
My workProduct concept · specs · workflow design · development with Codex and Claude Code
ExploreVisit distrove.app
01 / What I built
A workspace that goes from understanding a market to running the plan.
Distrove helps solo founders, agencies and service businesses move from understanding their market to preparing and executing a growth plan. It combines business analysis, market research, channel selection, prospect discovery, content creation and follow-up in one workspace, covering four roles: strategist, researcher, sales rep and copywriter.
02 / My role
My contribution.
I lead Distrove as founder and product owner.
- Product concept and direction, outcome definition and specifications.
- User journeys and workflow design across research, strategy, outreach and measurement.
- Development, implementation review and iteration, built hand in hand with Codex and Claude Code.
03 / How it works
Inside the system.
The browser manages the interface and calls server actions. The server validates requests, talks to providers and keeps state in SQLite. Background workers prepare recurring research, prospect batches, content and scheduled actions.
The weekly loop
- 01
Understand the business
Reads the website, records its sources and drafts a business profile. The founder confirms it and adds goals, customer value, buying process and past marketing attempts.
- 02
Research the market
Studies competitors, customer discussions and published growth stories from comparable businesses.
- 03
Build a strategy
Weighs channels against the business context and cited evidence, explains the recommendation, then a second AI review can challenge or revise the plan.
- 04
Prepare the work
Finds prospects, checks email addresses and writes personal outreach, follow-ups, posts and channel-specific material.
- 05
Review and execute
The founder approves every outward action. Connected workflows send or publish; other channels get prepared material for manual execution.
- 06
Track and learn
Records replies, meetings and reported outcomes, and keeps the learnings for the next review.
Technology stackThe building blocks
FrontendVanilla JavaScript ES modules · HTML · CSS
BackendNative Node.js server, no third-party runtime dependencies
DatabaseSQLite · account workspaces as JSON · records for jobs, subscriptions, usage and connections
AIOpenAI Responses API · structured output · image generation
ResearchExa search and page extraction · Hunter email discovery and verification · competitor-ad research
IdentityEmail and password accounts · server sessions · Google sign-in
IntegrationsX OAuth · SMTP / IMAP mailboxes · Resend · Stripe subscriptions
DeliveryHostinger · Node.js test runner with isolated fixtures and mocked providers
04 / Implementation
Under the hood.
The details that make the workflow dependable in practice.
Structured and checked
AI responses use structured output with server-side validation, and evidence references are checked before they reach the founder.
Workers that don’t collide
A database lease coordinates background work across several application instances, and provider failures have explicit handling.
A real mailbox workflow
Batch approval, paced delivery in sending windows, threaded follow-ups, reply matching with AI classification, pause controls and bounce handling.
Channels, honestly scoped
X publishing runs through OAuth. LinkedIn outreach is prepared for the founder to send by hand, and community research drafts replies for review.
Credentials encrypted
Stored connection credentials are encrypted, and sessions are kept on the server.
AI architecture
Where the model fits.
Models
OpenAI Responses API: a stronger model for strategy and writing, a lighter one for bulk extraction and classification, and image generation.
Output contract
Structured output, validated on the server before anything is stored or shown.
Evidence
Recommendations cite their sources and the references are checked. A shared library of researched public growth stories is reused across analyses.
Guardrails
Generated proposals remain subject to the founder’s permissions, approvals and spending limits.
Verification
Node.js test runner with isolated fixtures and mocked external providers.
05 / Why it works this way
Decisions with intent.
The thinking behind the architecture and the product experience.
Nothing goes out without a yes
The founder approves every outward action. Connected workflows send or publish only what was approved; other channels get material prepared for manual execution.
Evidence over business type
Channels are judged on the founder’s answers, cited market evidence and six criteria, then a separate skeptic pass can challenge or revise the plan.
Recommend first, then check what can run
What a business should do is decided separately from what the product can execute, so recommendations aren’t bent toward built-in features.
06 / The hard part
Recommendations that fit the business, not its category.
Early rules based on business type could push recommendations toward generic channels, including cold email for a self-serve product. I revised the strategy engine to weigh the founder’s answers, cited market evidence, six evaluation criteria and a separate skeptic pass. Recommendations now explain their basis, and a shared library of researched growth stories gives future analyses reusable evidence without repeating the extraction.
07 / Current stage
Where it stands.
Live product under active development. Business analysis, research, strategy, prospecting, content, the mailbox workflow and measurement are implemented.
270+Automated tests
0Runtime dependencies
9Provider integrations
EN/FRProduct and legal pages
Keep exploringNext projectViewMesh↗