AGORA — Minki Moon

mooner:~/resume/projects/agora $ tree -L 1

Structure⁰¹

구조

06 parts

A Ktor backend, a SvelteKit frontend and an extraction service for web pages. LLM calls go through an in-house client library that treats Claude, Gemini and OpenAI-compatible APIs the same way.

orchestration/: The orchestrator that runs an agenda, context assembly, the discussion event bus and per-role prompts.
nexus/: Web grounding: searches with SearXNG, fetches page text through the extractor and hands both to the agents as tools.
service/pdf: Extracts text from attached PDFs and estimates their token count up front.
extractor: An internal FastAPI service that turns a URL into Markdown and metadata with Trafilatura.
frontend: Writing agendas, watching the debate live, reports, per-role prompt and model settings, token usage stats.
auth: Passkey (WebAuthn) and TOTP sign-in.

mooner:~/resume/projects/agora $ cat docs/architecture.md

Architecture⁰²

아키텍처

Data flow

The debate runs on the server round by round, and every statement and tool call streams to the frontend over SSE. Agents call a web search tool when they need grounds.

Four roles: CTO, CCO/CMO, a decision manager, and a small model for summaries. Each role’s system prompt and model can be changed.
Each agenda sets its maximum rounds, a token budget and per-perspective weights, and can carry a PDF.
The frontend reads the fetch stream itself and renders statements, tool calls starting and finishing, and search results as they arrive.
The report holds the conclusion, confidence, key grounds, conditions, remaining risks and what would trigger a review.

mooner:~/resume/projects/agora $ git log --grep=fix

Problem solving⁰³

문제 해결

00 cases

// Being written.

Structure01

Architecture02

Problem solving03

Contact→

Structure⁰¹

Architecture⁰²

Problem solving⁰³

Contact^→