LLM Pulse?

Build a focused replacement for LLM Pulse, an AI search visibility monitoring platform. The core workflow: a marketing user creates a project, adds prompts about their brand, selects competitors, and schedules weekly monitoring runs. The system sends those prompts to multiple AI assistants (e.g., OpenAI ChatGPT, Perplexity, Google Gemini), stores the raw text answers, then analyzes each answer for: brand mention presence, mention position in the answer (first, second, etc.), cited URLs/domains, sentiment tone toward the brand, and competitor mentions. It aggregates results into the same metrics: Mention Rate (percentage of answers mentioning the brand), AI Visibility Score (position-weighted), Share of Voice vs each competitor, and citation overlap with competitors. A dashboard shows time-series trends, per-model breakdowns, sample answers, and content gap suggestions (topics where competitors appear but the brand does not). A content recommendation module (e.g., 'GEO Writer' only as a descriptive workflow term, not the original brand name) proposes briefs based on those gaps.

User-facing requirements: project/dashboard pages; prompt CRUD; competitor management; scheduling (weekly/daily); per-run status monitor; metric charts for mention rate, visibility score, share of voice, and sentiment over time; drill-down to individual raw AI answers and citations; simple report export (CSV/PDF).

Technical requirements: backend orchestrator that calls each AI provider's API asynchronously (use provider SDKs or REST), with retries and per-model rate limiting; a relational database (PostgreSQL) storing prompts, runs, raw answers, parsed mentions, citations, and sentiment; a queue (Redis + Celery or similar) for scheduled runs; an LLM-based extraction pipeline (e.g., a dedicated 'analyzer' model) that returns structured JSON for mention position/competitor/citations/sentiment; a lightweight REST API (FastAPI or similar) for the frontend; a Next.js or React frontend with charts (Recharts); local-only auth (no SSO) and SQLite fallback for demo mode.

Implementation approaches: use Pydantic schemas for extraction; store raw responses immutable; cache AI responses by prompt+model+date for idempotency; design for cost efficiency by batching prompts per model. Explicitly do not copy the original brand name, logo, visual design, DOM structure, or UI patterns; use a different name and a different visual theme. Out of scope: billing/subscriptions, multi-tenant enterprise features, native Rust CLI, SDK generators, Google Analytics integration, and paid model add-ons.

Deliver a README covering setup, required API keys, configuration of models, scheduler usage, database migrations, and important limitations: AI answers are non-deterministic, model APIs are rate-limited and costly, sentiment classification is approximate, and citation/footnote extraction varies by provider and may require per-provider parsers.