Moat

AI is the core, not a wrapper

Investor / advisor · 2026-05-10

In every AI module on the @appss stack, AI is the structural component, not a chatbot bolted onto a normal feature. Designer Studio orchestrates 14 tools across 5 providers. Market Research AI runs SQL on a live 121-table production database. Social Tracker uses a 3-tier cost-aware pipeline for 10× cost reduction. The orchestration is the moat, not any single model.

The pitch isn't «we use AI». Half the world uses AI. The pitch is that in every place AI shows up on our stack, the AI is the thing that makes the feature work, pull AI out and the feature collapses, not just degrades. That's a different category from chatbot-on-top-of-product, and it's hard to copy.

Where AI is core (not enabler)

Designer Studio

14 atomic tools across 5 external AI providers (Anthropic, OpenAI, recraft, replicate, quiver) under one orchestrator.
Claude Sonnet runs the multi-step pipeline: analyze → extract → vectorize → recurse → compose.
Multimodal: the agent SEES intermediate results and decides whether to recurse further, fall back to alternative paths, or ask the user.
Decision-making at the agent level: «is this trace clean enough?» → recurse / edit_image fallback / human prompt.
Not «AI generates an image», AI orchestrates a 14-tool brand-identity pipeline.

Market Research AI

Claude Sonnet 4 with GPT-4o fallback.
Tool-use over a live PostgreSQL database with 121 production tables and a PostHog event-stream with 60+ events.
Agent generates SQL on the fly, runs it, validates results, synthesises rich answer blocks.
Streaming SSE with rich response blocks rendered per-surface (web → full blocks; Telegram → TG-friendly via second parser-prompt).
«Never fabricate numbers» is enforced in the system prompt.
This is not chat-with-our-docs. This is an agent doing live data work on a real production schema.

Social Tracker

3-tier cost-aware pipeline: 1. Claude Haiku, caption screen (cheap) 2. Claude Haiku-vision, thumbnail check (cheap-ish) 3. Claude Sonnet + Whisper, deep video and audio analysis (expensive, gated)
Roughly 10× cost reduction vs. running everything through the heaviest tier.
Custom product-context system per app, same pipeline runs differently for a fitness app vs. a dating app.

Builder API

AsyncAnthropic, 5 analyzer endpoints.
Reads creators' GitHub repos → extracts SDK events / funnel / push / referral hooks.
Parallel-gather for multi-repo analysis.
This is the layer that makes cross-ecosystem porting feasible, same repo, multiple ecosystems.

Where AI is enabler (not core)

Pro dashboard, content recommendations, AI-powered insights.
Push management [planned], auto-generated copy for drip campaigns.
@appss search [planned], semantic search.

These are useful AI applications. They are not what we mean when we say «AI moat».

The full AI stack

Provider	Models we use	Used by
Anthropic Claude	Sonnet 4, Haiku, Haiku-vision	Designer Studio, Market Research AI, Social Tracker, Builder API
OpenAI	GPT-4o (fallback), gpt-image-1 (edit), Whisper	Market Research, Designer Studio, Social Tracker
recraft.ai	Vectorizer	Designer Studio
Replicate	rembg, [planned] SAM2	Designer Studio
quiver.ai	Text-to-vector	Designer Studio

Why this is a moat

Compositional moat. The individual providers are available to everyone. Our orchestration of all of them under a single Telegram-creator-flow is not. The cost of getting twelve capabilities to coordinate around one user-flow is paid in months, not weeks.
Data moat. Claude agents in Market Research read our 121-table production database, accumulated over 2+ years. A competitor with the same models and tools doesn't have the data layer. Same SQL, different answers.
Cost optimisation as engineering, not configuration. The 3-tier Social Tracker, recursion-stop heuristics in Designer Studio, prompt-caching across long-running agent sessions, these are engineered details, not flag flips. The race for $/operation is won by teams that polished for months.
Fallback resilience. Claude → GPT-4o; recraft → Replicate. We don't depend on one provider's uptime.

AI-driven UX patterns we use

Candidate model. AI generates a candidate-version; user hits «Apply» to commit. Never «AI silently changes things».
Multi-modal recursion. Agent observes the result of a tool-call and decides whether to dig deeper.
Style descriptor passing. One tool returns a structured-string ready for the next.
Fallback ladder. Step 1 → Step 2 → Step 3 (ask the user). Never «AI silently produced something different from what was requested».
Single answer, per-surface renderer. The canonical answer is produced once by the heavy model. Each UI surface (web, Telegram bot, future Discord / LINE) re-renders it via a cheap second parser-prompt (Gemini-Flash / Haiku) tailored to that surface's constraints. This is how cross-ecosystem expansion stays cheap, same answer, new renderer per surface.

Numbers and quotable claims

6 AI / ML providers integrated.
14 atomic tools in Designer Studio, 10 max iterations in agent loop.
3-tier pipeline in Social Tracker = ~10× cost reduction vs. naive heavy-model-everywhere.
121 production tables + 60+ event types queryable by Market Research AI.

Open questions

Total monthly AI spend.
Cost per user (Pro / Free / pay-as-you-go).
Fallback hit-rate (Claude → GPT-4o).
Average prompt-caching hit-rate.