Toolnaut — Role-Aware AI Tool Discovery & Learning

Toolnaut helps people cut through AI-tool overload. A 60-second quiz builds a career-aware persona, recommends a starter stack from a catalog of 700+ tools, and charts a 4-week learning roadmap with lessons and checkpoints. A background radar pipeline discovers, enriches, and publishes new AI tools automatically, so the catalog stays fresh with no human in the loop.

• Live: https://toolnaut.xyz
• Stack: React 19 · Vite · React Three Fiber (3D galaxy) · Framer Motion · Tailwind 4 · Node radar pipeline · Vercel

---

Architecture at a glance

Two systems share one catalog: the web app (what users see) and the radar pipeline (what keeps the catalog alive). They meet at public/tools.json.

flowchart TB
    subgraph Sources["Discovery sources"]
        GH[GitHub API]
        HN[Hacker News API]
        PH[Product Hunt API]
        RSS[RSS feeds]
    end

    subgraph Radar["Radar pipeline - Node"]
        direction TB
        F[filter: drop article/noise] --> D[dedup vs known.json]
        D --> E["enrich: categorize, blurb, tags<br/>NVIDIA LLM or rules"]
        E --> V[validate + score]
        V --> C[courseGen: lessons + skills]
        C --> ST{confident enough?}
    end

    subgraph DB["Store - JSON DB"]
        PUB[(published tools.json)]
        REV[(review-queue)]
        CRS[(courses / skills)]
    end

    subgraph App["Web app - React/Vite"]
        direction TB
        BUNDLE[Bundled catalog: 704 tools] --> HYD[liveCatalog hydrate at startup]
        LIVE[public tools.json] --> HYD
        HYD --> QUIZ[9-question quiz]
        QUIZ --> PERSONA[personaGenerator: career-aware persona]
        PERSONA --> STACK[Recommended starter stack]
        STACK --> ROADMAP[4-week roadmap: lessons + checkpoints]
    end

    Sources --> F
    ST -->|yes| PUB
    ST -->|no| REV
    C --> CRS
    PUB -->|npm run sync| LIVE
    App -->|vercel deploy| VERCEL[Vercel production]

The autonomous loop: sources → filter → dedup → enrich → validate → course → publish → sync → deploy → the app hydrates the new tool → the quiz recommends it.

---

The web app

Route	Purpose
/	3D galaxy landing (hero, capabilities, contact)
/quiz → /quiz/result	9-question quiz → career-aware persona + starter stack
/app/stack · /discover · /tools/:slug	Your tools · browse the catalog · tool detail
/app/learning	4-week roadmap: expandable lessons + gating checkpoints
/app/community · /settings	Threads · preferences

How a recommendation is made

1. quizLogic.js — 9 questions capture domain, role, career stage, experience, goal, budget, pace, learning style, blocker.
2. personaGenerator.js — turns answers into a persona (e.g. Ambitious Designer · Senior Designer), picks the top-3 starter stack from the catalog, and suggests a plan tier (Student / Pro / Team).
3. roadmapGenerator.js — builds a 4-week roadmap; pace scales the number of steps, each step carries a lesson, and each week ends with a checkpoint quiz that gates the next week.

Catalog hydration — liveCatalog.js fetches /tools.json (the radar's output) at startup and merges any new tools over the bundled 704. A missing file is a safe no-op, so the app always renders.

Play modes — a theme switcher (themeStore.js) swaps the accent palette (Nebula / Solar / Toxic / Synth) live via CSS variables.

---

The radar pipeline (radar/)

A standalone Node pipeline. Runs with zero keys (public sources + rule-based enrichment → review queue); keys unlock AI enrichment + auto-publish.

radar/
├── run.js            # entry: loads env, runs the pipeline, logs a summary
├── env.js            # auto-loads radar/.env (gitignored)
├── config.js         # all config from env, with safe defaults
├── sources/          # github · hackernews · producthunt · rss connectors
├── filter.js         # drops article/headline noise (is-it-a-tool)
├── dedup.js          # skips tools already in known.json
├── enrich.js         # categorize · blurb · tags (LLM or rules)
├── llm.js            # provider-agnostic: Anthropic -> NVIDIA -> OpenAI -> OpenRouter
├── validate.js       # schema + confidence scoring
├── courseGen.js      # generates lessons + skills per tool
├── store/            # JSON DB (published, review-queue, courses, known)
└── scripts/sync-to-app.js  # published tools -> ../public/tools.json

Run it

cd radar
cp .env.example .env      # add keys (optional — see below)
npm test                  # pipeline tests
node run.js               # hunt -> enrich -> store
npm run sync              # push published tools to the app
node run.js && npm run sync   # full cycle

Keys (radar/.env, gitignored — never committed):

Key	Effect
NVIDIA_API_KEY	AI enrichment + auto-publish (biggest power-up)
GITHUB_TOKEN	higher GitHub hunt rate limit
PRODUCTHUNT_TOKEN	enables the Product Hunt source

LLM provider is auto-selected (Anthropic → NVIDIA → OpenAI → OpenRouter); the first key present wins. NVIDIA uses meta/llama-3.1-70b-instruct by default.

---

Run & deploy the app

npm install
npm run dev      # http://localhost:3000
npm run build    # production build -> dist/
vercel deploy --prod --yes   # deploy to Vercel (project already linked)

The app is a client-rendered SPA. vercel.json rewrites all routes to index.html. Deploys are linked to the Vercel project nexus.

The brand name lives in one place — src/config.js (BRAND).

---

Status

Beta — web-first. The discovery → recommend loop is live and autonomous. Known follow-ups: SSR/prerender for SEO crawlability, a mobile navigation menu, and scheduling the radar (cron / GitHub Action) for daily unattended runs.