ComradeIQ

An AI command center where a Commander AI orchestrates specialized AI comrades to plan, execute, and merge complex tasks into one intelligent, high-quality result.

Comment

One prompt. A whole team. A finished deliverable — and you watch every step happen.

ComradeIQ is an AI mission control: you give Commander Atlas an objective, and a coordinated team of specialists — Researcher, Writer, Critic, Formatter, and Assembler — works through a real dependency pipeline to hand you back a verified, downloadable result. Not another chat box. A command center.

💡 Inspiration

Most "multi-agent" demos are theater — spinning cards that invent their own progress, phantom parallel calls, and confident answers with no traceable source. We wanted the opposite.

ComradeIQ started from a single question:

What would AI coordination look like if every visible step had to be **provably true?

No simulated research. No fabricated artifacts. No fake agents. Just a Commander with an honest dependency graph and a team that only activates when it's genuinely needed.

🎯 What it does

You type one objective. ComradeIQ routes it to the right workflow automatically:

  • 💬 Direct chat — quick questions get a fast, single answer, no overhead.
  • 📄 Document missions — ask for a README, spec, or report → get a downloadable Markdown file.
  • 📊 Presentation missions — ask for a deck → get a styled, downloadable PPTX in one of 4 themes (Camo, Cyberpunk, Minimal, Ocean).
  • 🔬 Research missions — opt-in web access returns sourced, cited answers — never invented URLs.

For anything that needs coordination, the Commander builds an explicit dependency DAG and executes it for real:

Researcher → Writer → Formatter → Critic → Assembler → Commander QA

Each specialist waits for its actual upstream output. And you see it happen two ways at once:

  • 🟢 Live agent graph — every specialist node lights up in real time (thinking → working → done) as work flows down the graph.
  • 🖥️ Live agent console — a terminal-style ops log streaming the real mission events (dispatch --agent writer, [critic] working, ✓ mission complete), with each line expandable to reveal that agent's actual contribution.

And two things purely for delight:

  • ♟️ Play chess with the Commander — say "play chess with me" and a real, rules-validated board opens in chat. You're White; Commander Atlas plays Black through the same LLM that runs missions — literally the Commander playing you.
  • ▶️ In-chat video — ask for "a video of how jet engines work" and ComradeIQ searches YouTube and embeds a playable player right in the conversation.

Finished missions get shareable read-only permalinks, and every chat has a hover menu to share, archive, or delete it.

If a provider, storage layer, or capability isn't configured, ComradeIQ says exactly what's missing — it never fakes a capability.

🛠️ How we built it

  • Next.js 15 + React 19 + TypeScript — a responsive chat workspace with server-only API routes and end-to-end type safety.
  • Groq (llama-3.3-70b) over an OpenAI-compatible API — provider-agnostic model calls with structured routing, JSON-Schema output for decks, and vision-ready image input.
  • A dependency-aware orchestrator — a real DAG, not cosmetic parallelism. Upstream results are the literal inputs to downstream agents, and each role sees only its own inputs — no shared transcript.
  • SSE-first streaming — the plan, dispatch, per-agent status, and result all stream to the browser with no external realtime dependency.
  • PptxGenJS for server-side decks, chess.js for a legality-validated board with an LLM opponent, and a keyless YouTube search + embed for in-chat video.
  • Zustand for client state, Vercel Blob for durable mission/artifact storage, and Vitest + Playwright for routing, DAG, state, a11y, and reconnect coverage.

🧗 Challenges we ran into

Resisting theatrical shortcuts. Making a multi-agent UI look impressive is easy; making every visible update trace to a real event, every artifact to real storage, and every error to an honest error took discipline at every layer.

Serverless reality. The deployed build 500'd on every API route — but only in production. The culprit: the production bundler skipped Node file tracing, so a heavy dependency (pdf-parse) crashed each function at import time. We diagnosed it from the raw response headers (X-Matched-Path: /500, no platform error), moved the build to the stable path, externalized the offending packages, and lazy-loaded them off the request path.

Honest state everywhere. Ownership boundaries, cancellation, reconnect, and truthful "what's missing" states turned out to be as much work as the model calls themselves.

🏆 Accomplishments we're proud of

  • A real mission dependency DAG — not decorative agent cards.
  • A live agent graph + ops console that visualize the actual pipeline, with click-to-expand per-agent output.
  • Play chess against the Commander and in-chat video search & embed — real, interactive, embedded.
  • Downloadable Markdown & PPTX artifacts, 4 deck themes, and shareable permalinks.
  • A product that never fabricates — honest providers, storage, and capabilities throughout.

📚 What we learned

Reliable AI UX is mostly systems design — routing, observability, cancellation, and truthful errors matter as much as the prompt.

We also learned why decomposition wins. A single model gives you one pass. Add a Critic that catches a fraction $p$ of a draft's errors, and the expected residual error becomes:

$$\varepsilon_{\text{final}} \approx \varepsilon_{\text{draft}}\,(1 - p)$$

The innovation isn't a new model — it's role specialization + a dependency graph + an explicit review stage, made visible and shippable. That's how real production AI is built.

🚀 What's next for ComradeIQ

Authenticated team workspaces, richer source-provenance UI, evaluation dashboards that compare artifact quality across routing strategies, connector-backed business data, and configurable retention — kept honestly as future work rather than pretending they already exist.

Built With

Share this project:

Updates