Kobopilot

Interactive trading education platform — learn, quiz, and practice with paper trading.

Features

• Landing page — marketing site at /
• Dashboard — progress overview at /dashboard
• Learn — 8-lesson path with quizzes at /learn, plus a glossary at /learn/glossary
• Practice — paper trading simulator at /practice
• Kobo AI — an NGX coach in every screen, grounded in live quotes, your paper portfolio, and your lesson progress

Getting started

npm install
cp .env.example .env.local   # optional, see below
npm run dev

Open http://localhost:3000.

Environment

Everything is optional — the app runs with no configuration at all.

Variable	Effect if unset
OPENAI_API_KEY	Kobo AI falls back to scripted offline replies
OPENAI_CHAT_MODEL	Defaults to gpt-5-mini
OPENAI_REASONING_EFFORT	Defaults to minimal (gpt-5* models only)
NGXPULSE_API_KEY, TROVE_API_KEY, ITICK_API_TOKEN	Prices are deterministically simulated

Market providers are tried in order and the first returning 3+ quotes wins. Leaving a variable blank counts as unset.

Market data

NGXPulse is the working provider. It gives a snapshot of all NGX equities and daily closing bars per symbol — no intraday, no published intraday high/low, no websocket. Equities refresh about every 20 minutes and the exchange trades 09:00–16:00 WAT, Mon–Fri.

Two things follow from that, and both are load-bearing:

Quota. The Personal tier allows 10 requests/min and 100/day. Client polling is therefore decoupled from upstream calls: the snapshot fetch carries revalidate: 1200 and history revalidate: 86400, so one upstream call serves every connected client, and lib/market/market-hours.ts stops the client polling altogether while the NGX is shut. Worst case is roughly 30 upstream calls a day. Public holidays are not hardcoded — they are detected from the feed's own trade_date going stale.

The chart has two modes. ngx live draws real daily closes over 1w/1m/3m/1y. Because real prices barely move within a session, demo mode draws a simulated intraday series so learners can see what price action looks like. Each mode states which it is on screen, and synthesis never runs under the live label — getNgxHistory deliberately has no simulated fallback, because serving invented history from a "live" chart would present fabricated data as real.

Kobo AI

POST /api/chat streams replies via the Vercel AI SDK. The model answers through tools rather than from memory: get_quote and get_watchlist hit the same NGX provider chain the charts use, while get_portfolio and get_progress read a snapshot the client sends with each request, since that state lives in localStorage.

Guardrails worth knowing about before changing the prompt or tools:

• get_lesson never returns the quiz field — the model cannot leak quiz answers.
• Prior tool-call parts are stripped from the transcript server-side, so a crafted request cannot feed the model fabricated tool output.
• The endpoint is unauthenticated, so it is rate limited to 20 requests per 10 minutes per IP, held in process memory. Replace this with a Redis-backed limiter before a public launch — see lib/ai/rate-limit.ts.

Tech stack

• Next.js App Router
• TypeScript
• Tailwind CSS + shadcn/ui
• Zustand (localStorage persistence)
• Vercel AI SDK + OpenAI

Brand

Kobopilot uses the Invest Bamboo design token system (deep green, warm canvas, lime accents). See .cursor/skills/kobopilot-design/SKILL.md and .cursor/skills/investbamboo-design/SKILL.md.