● refactoringbuildingsamridhlimbu.com/projects/hoddle · v0.1

❯ cd projects/hoddle

Hoddle Melbourne

● live · CTO

International mentorship platform connecting first-year students at Melbourne universities with verified senior mentors. Built end-to-end as sole developer — schema, auth, matching algorithm, community layer, and background infrastructure.

hoddle.org code

10+

verified mentors

100+

early-access mentees

14,040

linkedin impressions

Context

CTO and sole developer, Mar 2026–present. Hoddle is a one-to-many mentorship platform — mentors publish articles, host live Q&As, and tell their story; students get matched based on shared background and goals, browse a content library, and ask questions in forums. Eventually students graduate into mentors themselves.

I owned the full stack from day one: database schema and RLS policies, auth flows, matching algorithm, community features, background jobs, and admin tooling. The codebase sits at 26,000+ lines across 51 routes, 38 components, 19 server action files, and 17 migrations.

Built as part of the Study Melbourne Leadership 4Impact Program alongside co-founders Aarav Verma and Tanvir Kaur Sohi. The team placed 2nd at the final pitch day.

Timeline

Mar 2026

Foundation

Joined as CTO and sole developer. Rejected every third-party auth provider — one Supabase service for auth, Postgres, and RLS. Set up full project scaffold: design tokens, Tailwind v4 theme, route groups, Supabase clients (server, browser, admin), and TypeScript strict mode.

Mar 2026

Auth & onboarding

Email magic link auth via admin.auth.admin.generateLink() and nodemailer, bypassing Supabase SMTP entirely. 5-step student wizard — university, country of origin, goals, challenges, fields of interest — Zod-validated on both client and server. Partial saves persist between steps.

Apr 2026

Matching algorithm

scoreMentor(): country +30, field overlap +15 each, challenges and goals +10 each. IO-separated — zero database calls in the scoring function. Scores stored in mentor_recommendations, recomputed nightly at 02:00 AEST via Vercel cron and on-demand after onboarding updates. "Why this mentor" reasoning line rendered on the dashboard.

Apr 2026

Mentor invite system

Token-gated invite flow: admin sends an email generating a row in mentor_invites with a hashed token, 14-day expiry, and audit columns. Mentor clicks through to /mentor-signup/[token], validated server-side before rendering. After magic-link auth, acceptMentorInvite() sets role=mentor and routes to a 4-step onboarding wizard.

Apr 2026

Content library

Mentors author articles with a Tiptap rich-text editor, embed YouTube/Vimeo via URL, and attach downloadable resources stored in Supabase Storage with signed URLs (25 MB limit). Students browse with type filter chips and offset pagination (18 per page). View counts increment atomically via server action on each load.

Apr 2026

Community forums

5 seeded categories: First Semester Struggles, Career & Internships, Living in Melbourne, Academic Writing, Visa & Admin. Threads with two-level nested replies, 4 reaction types (Helpful, Thanks, Hugs, Insightful), view counter via SECURITY DEFINER RPC, anonymous posting for students, quote-reply and @mention shortcuts. Mentors always post as identified.

Apr 2026

Live Q&A and sessions

Mentors schedule sessions with title, description, datetime, duration, capacity, and a meeting URL. Students register (capacity-enforced server-side), submit questions with an anonymous toggle, and receive recordings after. Three Vercel cron jobs handle the full session lifecycle.

Apr 2026

Notification system

Single notify(recipientId, type, payload) helper writes to the notifications table and conditionally dispatches email based on per-type, per-channel preference toggles. In-app bell uses a Supabase Realtime subscription — unread count stays live without polling. 6 notification types.

Apr 2026

Direct messaging

Private conversations between students and mentors: conversations, messages, and conversation_read_cursors tables. Rate-limited to 30 messages per 10 minutes. new_chat_message notification fires with a 5-minute per-conversation debounce to avoid notification flooding.

Apr 2026

Google OAuth + feedback pipeline

Added Google sign-in alongside magic link. Floating feedback widget on all browse pages (Bug / Suggestion / Confusion / Other) forwarded to Airtable via a server action. Auth flows through /auth/confirm client component that handles the implicit-flow hash fragment the server callback route cannot reach.

Key technical decisions

supabase rls › application-layer guards

Row Level Security enforces data access at the database layer. A bug in middleware can expose data; a bug in application code cannot bypass RLS. Every table has RLS enabled with explicit policies. The admin client is the only surface that bypasses RLS — used only for system operations like notify() and recommendation recompute.

three supabase clients › single shared client

Server client is cookie-bound, used in RSCs and server actions. Browser client powers the Realtime notification bell. Admin client uses the service role key and is never imported from app/ or components/ — the key has no NEXT_PUBLIC_ prefix and never ships to the browser.

token table for invites › magic link via email provider

Owning the invite row means owning the full lifecycle: custom expiry, re-invite without blocking, revocation, reuse prevention, and full audit trail via created_at, expires_at, accepted_at. No dependency on third-party link delivery semantics.

pure scoring function › coupled db + score logic

scoreMentor() takes two plain objects and returns a score and a reasoning string — zero database calls. computeRecommendationsForProfile() handles all Supabase interaction separately. The scoring function is importable into tests with no mocking required.

nightly precompute › on-demand scoring

Match scores are read far more often than profiles change. Precomputing nightly keeps the dashboard to a single ORDER BY score DESC query, running at 02:00 AEST. Scores are also recomputed immediately after any profile update — on-demand only when stale.

nodemailer + gmail smtp › supabase smtp

Generates the magic link server-side via admin.auth.admin.generateLink() and sends via nodemailer. Removes Supabase SMTP config dependency and fixes a production bug where links landed at /?code=… because NEXT_PUBLIC_SITE_URL was missing from Vercel env.

claude.md + arch docs › ad-hoc onboarding

The repo has a CLAUDE.md master briefing, docs/architecture.md, docs/database-schema.md, docs/component-library.md, and docs/design.md. Future contributors and agents can understand the system without a walkthrough. Claude Code uses these docs on every task before touching a file.

The matching algorithm

lib/matching/score.tsts

1export function scoreMentor(student, mentor): ScoreResult {
2  let score = 0
3  const reasonParts: string[] = []
4
5  // country of origin — strongest signal (+30)
6  if (student.country_of_origin === mentor.country_of_origin) {
7    score += 30
8    reasonParts.push(`Also from ${mentor.country_of_origin}`)
9  }
10
11  // field of interest overlap (+15 each)
12  const fieldMatches = student.fields_of_interest.filter(
13    f => mentor.expertise.includes(f)
14  )
15  score += fieldMatches.length * 15
16
17  // challenges + goals overlap (+10 each)
18  score += student.challenges.filter(c => mentor.expertise.includes(c)).length * 10
19  score += student.goals.filter(g => mentor.expertise.includes(g)).length * 10
20
21  return {
22    score,
23    reasoning: reasonParts.join(" · ") || "Verified mentor on Hoddle",
24  }
25}

Country is weighted highest because shared migration context — knowing what it's like to arrive from the same country — matters more than overlapping in the same field. The reasoning string is stored alongside the score and rendered on the student dashboard as a "why this mentor" line. Scores above a threshold of 10 count as real matches; below that, the algorithm backfills from top-ranked verified mentors to always surface five recommendations.

Database

17 migrations · ~1,500 lines of SQL

profilesonboarding_responsesmentorsmentor_invitescontent_itemscontent_resourcescontent_tagsforum_categoriesforum_threadsforum_postsforum_reactionssuccess_storieslive_sessionssession_registrationssession_questionsnotificationsnotification_preferencesmentor_recommendationsmentor_followsconversationsmessagesconversation_read_cursorsreviews

storage buckets

content-media (public)content-resources (signed)session-recordings (private)story-images (public)avatars (public)

Background jobs

/api/cron/session-reminders

24h pre-session reminder emails to all registrants

hourly

/api/cron/session-starting-soon

In-app + email push 10–20 min before sessions start

every 5 min

/api/cron/recompute-recommendations

Recompute match scores for all onboarded students

nightly 02:00 AEST

Stack

FrontendNext.js 16 · TypeScript · Tailwind v4

BackendSupabase · PostgreSQL · RLS

InfraVercel

Scale

26K+

lines

routes

components

server actions

migrations

“They had a really easy time navigating through the UI — the experience matched the promise.”

Early-access cohort · Hoddle Melbourne

What's next

Phase 3 starts after two weeks of production telemetry. The platform is designed to extend — every feature boundary is RLS-enforced and IO-separated so new surfaces can be added without touching existing data contracts.

›

Mentor badges + achievement tiersgamified progression for verified mentors

›

Mentee → mentor graduation programstructured pathway for students to give back to the next cohort

›

Multi-language UIMandarin, Hindi, Vietnamese, Korean — the four largest student cohorts

›

PWAinstallable shell, offline mode, web push notifications

›

University calendar integrationsUnimelb, Monash, RMIT ICS feeds synced to session scheduling

›

Matching algorithm v2telemetry-driven collaborative filter replacing the static weight model

›

AWS backend migrationLambda + RDS for multi-region deployment, reduced cold start latency, and decoupling from Vercel function limits

hoddle.org source ← back to projects

1export function scoreMentor(student, mentor): ScoreResult {

2 let score = 0

3 const reasonParts: string[] = []

5 // country of origin — strongest signal (+30)

6 if (student.country_of_origin === mentor.country_of_origin) {

7 score += 30

8 reasonParts.push(`Also from ${mentor.country_of_origin}`)

9 }

11 // field of interest overlap (+15 each)

12 const fieldMatches = student.fields_of_interest.filter(

13 f => mentor.expertise.includes(f)

14 )

15 score += fieldMatches.length * 15

17 // challenges + goals overlap (+10 each)

18 score += student.challenges.filter(c => mentor.expertise.includes(c)).length * 10

19 score += student.goals.filter(g => mentor.expertise.includes(g)).length * 10

21 return {

22 score,

23 reasoning: reasonParts.join(" · ") || "Verified mentor on Hoddle",

24 }

25}