Polaris — AI Academic Strategist

The Infinity AI BuildFest 2026 submission. A fully autonomous AI academic strategist that reverse-engineers competitive university-admission profiles for students — from early schooling to graduate school.

Polaris is an AI-native platform that replaces fragmented Reddit threads and $2,000+ consultants with a transparent, data-driven roadmap.

What's inside

Layer	Stack
Frontend	Next.js 15 (App Router) · TypeScript (strict) · Tailwind CSS
Auth	NextAuth v4 (JWT) · credentials · bcrypt · MongoDB Atlas
Roles & plans	RBAC (student / parent / partner / admin) · Free / Pro / Elite tiers with server + client feature gating
Payments	LemonSqueezy (merchant-of-record) — hosted checkout, customer portal, signed webhooks
LLM	Google Gemini 2.0 Flash with structured-JSON output (responseSchema)
Retrieval	RAG over a curated 60-doc KB (universities, scholarships, accepted-student case studies) using Gemini embeddings + in-memory cosine search
ML	Logistic-regression acceptance-probability model with transparent factor weights — no demographic proxies
Hardening	Zod validation · centralized API error handling · MongoDB indexes · env validation · security headers
Localization	English / বাংলা
Fallback	Deterministic heuristic engine — the demo works even with no API key

Demo flow (the vertical slice)

1. Landing (/) — problem framing, 3-tier pricing, AI-architecture chips.
2. Intake (/onboard) — 6 questions captures grade, country, degree, GPA, ECs, target tier.
3. AI Strategist (/api/roadmap) — RAG over the KB + Gemini structured-JSON returns 8–12 milestones grouped by quarter, with rationale grounded in real KB docs.
4. Dashboard (/dashboard) — strategic summary, "what successful applicants had that you don't yet", quarter-grouped milestones, next-30-days focus card.
5. Probability Engine (/university + /university/[id]) — picks a target university from a curated list of 20, scores acceptance probability, shows the top-contributing factors with weights, and lets you simulate "what if I add X" with live sliders.

Run locally

npm install
cp .env.local.example .env.local
# (optional but recommended) drop your free Gemini key into .env.local
npm run embeddings   # one-off: builds data/embeddings.json (skips silently with no key — keyword fallback kicks in)
npm run dev          # http://localhost:3000

Without a Gemini key Polaris transparently degrades to (a) keyword-overlap retrieval and (b) the heuristic roadmap engine — so the full demo flow is verifiable offline.

Architecture decisions worth flagging for judges

• Structured-output LLM tool-use — not a chatbot wrapper. Gemini returns typed JSON conforming to a responseSchema, so the dashboard can render confidently.
• No vector DB — ~60 docs × 768 dims live in data/embeddings.json; cosine search runs in pure TS. Zero infra, deterministic, fast enough.
• Transparent ML — every probability shows top-3 contributing factors and their weights. Ethical-AI by design — no demographic features.
• Modular RAG — adding a country = adding a JSON doc + re-running npm run embeddings. The architecture scales globally without code changes.

Accounts, plans & B2B2C

Polaris is account-based with role-aware access:

• Roles — student (default), parent, partner, admin. Chosen at sign-up; parents/partners get a read-only monitoring dashboard (/monitor).
• Plans — free / pro / elite, enforced both server-side (requirePlan in lib/authz.ts) and client-side (usePlan / canUse in lib/features.ts). The feature→tier map lives in lib/features.ts.
• Parent/partner linking — a student invites someone by email from /family; the invitee accepts via a token and then sees progress (read-only) at /monitor.

Feature gating

Feature	Min plan
University/scholarship directory, acceptance rates, probability teaser, case-study titles	Free
AI roadmap generation, milestone tracking, benchmarking, full case studies	Pro
Adaptive replanning, professor lists, analytics, strategy reports, scenario sims	Elite

Payments (LemonSqueezy)

Subscriptions are handled by LemonSqueezy (works for Bangladesh-based payouts as a merchant-of-record):

1. Create a store + two subscription products ("Polaris Pro", "Polaris Elite") at lemonsqueezy.com.
2. Copy each variant id and create an API key.
3. Add a webhook → <domain>/api/webhooks/lemonsqueezy (all subscription_* events) and copy the signing secret.
4. Fill the LEMONSQUEEZY_* vars in .env.local (see .env.local.example).

Checkout (/api/checkout) launches a hosted checkout; the signed webhook (/api/webhooks/lemonsqueezy) flips the user's plan; /billing opens the customer portal. With LemonSqueezy unset, paid CTAs degrade gracefully and the Free tier still works end-to-end.

Built for

The Infinity AI BuildFest 2026.