AdMate Insights
Build "AdMate" — an AI marketing analyst for TikTok Ads.
Core concept: AdMate tells marketers (1) what changed, (2) what actually needs attention, (3) possible explanation, (4) recommended action, (5) evidence, (6) confidence.
Final flow: TikTok Ads data → AdMate analysis → recommendation → Telegram briefing.
Important architectural rule: strictly separate deterministic computation from AI.
- Code (not AI) calculates metrics and classifies findings (anomaly detection, trend breaks, threshold crossings, severity).
- AI is used only to explain findings and write recommendations/confidence framing in natural language, based on the structured output the code already produced. AI must never be the source of the numbers or the classification itself.
Architecture to implement:
DemoTikTokDataProvider — generates realistic synthetic TikTok Ads data (campaigns, ad sets, ads, daily performance metrics: spend, impressions, clicks, CTR, CPC, CPA, conversions, ROAS, frequency, etc.) over a multi-week period, with some intentional anomalies/trends baked in (e.g. a CPA spike, a CTR drop, a creative fatigue pattern, a budget pacing issue) so the analysis engine has real things to detect. Build it behind a clean data provider interface/contract so it can later be swapped for a LiveTikTokDataProvider (real TikTok Ads API) without changing any downstream code. No CSV upload — data is generated internally.- A normalization layer that maps whatever the provider returns into one consistent internal schema, so the rest of the app never depends on the provider's raw shape.
- An analysis engine (deterministic, code-based) that processes normalized data and produces structured findings: what changed, severity/what needs attention, supporting evidence (the actual numbers/comparisons), all computed and classified in code — no AI involved in this step.
- An AI explanation layer that takes each structured finding and generates: a possible explanation, a recommended action, and a confidence level/reasoning, in natural language. This should use Lovable AI. Make clear in the code/comments that this layer only explains — it does not compute or classify.
- A recommendation UI/dashboard where marketers can see each finding as a card/entry showing: what changed, why it needs attention, possible explanation, recommended action, evidence, and confidence — clearly organized and easy to scan, prioritized by severity.
- A Telegram briefing feature: send a digest of the top findings and recommendations to a Telegram chat via a Telegram bot. This should be triggerable manually from the UI for now (e.g. a "Send briefing to Telegram" button), producing a well-formatted summary message.
Please set up Lovable Cloud for the database (storing synthetic data, findings, analysis history) and edge functions (for running the analysis engine and calling the AI + Telegram). Use Lovable AI for the explanation/recommendation layer. For Telegram, use a bot token stored as a secret and ask me for it when needed.
Design: clean, professional analytics/dashboard aesthetic — this is a tool for marketers making decisions, so prioritize clarity and scannability over decoration.
This project was built with Lovable.
Build with Lovable
Continue developing this project in the Lovable editor.
- Ship faster: describe what you want to build and Lovable handles the code.
- Stay in sync: every change made in Lovable is committed straight to this repository.
- Full ownership: this code is yours. Push to
main on GitHub and your changes sync back into Lovable, ready for your next prompt.
Development
Prefer working locally? You need Node.js and npm — install with nvm.
git clone <this-repository-url>
cd <repository-name>
npm i
npm run dev
