Founder @ Mirembe Muse · Built 7 AI-powered SaaS apps in 9 months · Next.js · TypeScript · Claude API · Supabase

github

Nandawula Kabali-Kagwa • Peerlist

True Access - Accessibility & Location-Based Service PWA Role: Technical Founder | Full Stack Developer Year: 2025 Technologies: HTML, CSS, JavaScript, Supabase (PostgreSQL), Vercel, GitHub 📌 Overview True Access is a PWA app designed to help people with disabilities easily find verified, accessible locations and accommodations across South Africa. The platform provides a simplified search interface, detailed accessibility profiles, and a scalable database structure, solving a real gap in accessible travel, safety, and mobility information. 🛠️ Key Responsibilities ✔ Full Frontend Development: • Built responsive UI with clean HTML, CSS, and JavaScript • Designed search and filter system for quick data access • Implemented user-friendly listing cards with accessibility details •users can also leave detailed reviews and vote on other users reviews. ✔ Backend Architecture (Supabase): • Structured the PostgreSQL database for 5,000+ future listings • Created tables for locations, accessibility features, user roles, Admin roles  and user data • Configured secure Mapbox API access • Implemented scalable real-time fetch logic • Deployed using Vercel • Integrated CI/CD workflows • 40+ Git contributions during development • Set up environment variables and project configs 📂 Features Included 🔎 Smart Search & Listing System • Search bar on the Mapbox Interface • Filtered map pins for Accomodation, retail, gov. etc • Sorting based on accessibility criteria 🏨 Location Profiles Every listing includes: • Name & type of location • Address & Ratings • Wheelchair accessibility • Ramp availability • Parking details • Assistive features 📊 Supabase Data Structure • Normalized tables • Easy CRUD management • Structured for future dashboards 📱 Mobile Responsive Design • Seamless UX on phones, tablets, and desktop • Optimized layout for accessibility For People With Disabilities: • Easier access to verified, safe, accessible places • Supports independence and mobility

True Access

Built a facility booking platform featuring real-time synchronization, AI assistance, and WhatsApp integration (still on-going project that can be improved). 🎯 Highlights: - Full-stack Next.js 14 + TypeScript application with serverless architecture - Real-time updates using Supabase Realtime WebSocket subscriptions - Secure authentication with Row Level Security and role-based access - AI booking assistant powered by OpenAI API - WhatsApp Cloud API for conversational bookings - QR code generation for facility entry passes - Admin dashboard with live facility monitoring 💻 Stack: Next.js 14, TypeScript, React 18, Supabase (PostgreSQL + Auth + Storage + Realtime), TailwindCSS, WhatsApp Cloud API, OpenAI API 🔧 Advanced Features: ✓ Database schema with foreign keys and cascading deletes ✓ Row Level Security policies for data isolation ✓ Real-time WebSocket subscriptions with cleanup ✓ Webhook handler for WhatsApp commands ✓ Type-safe API routes with validation ✓ Conflict prevention via unique constraints ✓ Equipment checklist system (22+ items) 📊 Scale: 40+ components, 8 API routes, 3 database tables, 6 RLS policies, 4 API integrations, 2,000+ lines of code, 8 facilities This project demonstrates my ability to architect and deploy complex full-stack applications with modern technologies, external APIs, and production-grade security.

Cortex Hub Booking System – Full-Stack Facili

ADMINOS — AI Business OS for SA SMEsMulti-tenant AI OS for South African SMEs, NGOs, schools, and clinics. Built to replace six separate subscriptions with one WhatsApp-native platform that actually understands how African businesses operate.<h3>Architecture</h3>The core is a 7-step Workflow Engine: load tenant context → classify intent → check FAQ cache (Upstash Redis, 7-day TTL) → generate Claude response with prompt caching → deliver via WhatsApp → write to immutable audit log → push to dashboard via Supabase Realtime. Full pipeline resolves in under 3 seconds.Multi-tenancy is enforced at the Supabase RLS layer, not application code. Every table carries tenant_id verified against auth.jwt(). A bug in the app cannot leak one business's data to another by architectural design.<h3>Five Specialist AI Agents</h3>Inbox Agent — handles WhatsApp conversation routing and response. Classifies every inbound message into one of 12 intent categories before routing to the correct agent. Maintains conversation context per contact across sessions using Redis.Debt Recovery Agent — 5-tier automated escalation over 30 days. Claude drafts each message in the tenant's own tone: reminder → firm reminder → formal notice → legal warning → final demand. Escalation is cron-driven (Vercel Edge Cron), fires at 05:00 SAST daily, reads outstanding invoices from Xero API, generates personalised messages per debtor.Wellness Agent — staff check-ins via WhatsApp on a configurable schedule. Tracks response sentiment over time. Flags to the manager dashboard when a team member's average wellness score drops below 2.5/5 for 7 consecutive days. Built because SA SME owners often have no HR infrastructure but still lose good people to burnout they never saw coming.Document Intelligence Agent — accepts PDF, DOCX, XLSX via WhatsApp. Claude classifies the document type, extracts key fields, routes to correct table in Supabase, and returns a structured summary to the sender. Built for clinics processing patient referral letters and NGOs managing grant compliance documents.Analytics Agent — generates a Daily Manager Brief delivered to the owner's WhatsApp at 07:00 SAST every weekday. Covers: overnight messages handled, outstanding invoices, staff wellness average, document processing queue, and one AI-generated observation about the previous day's patterns.<h3>Cost Engineering</h3>Prompt caching on the static system context saves 40–60% on Claude API costs per request. Haiku routing handles structured, deterministic tasks (intent classification, document field extraction). Sonnet handles nuanced, emotionally intelligent responses (wellness conversations, client-facing debt messages). The routing decision is made per-request based on a complexity score derived from message length, intent classification confidence, and conversation history depth.<h3>Payment and Compliance</h3>PayFast integration for ZAR billing. Per-tenant rate limiting via Upstash sliding window (100 requests/minute). POPIA-compliant data handling — Mirembe Muse POPIA Reg. No. 2026-005658. Xero API integration for invoicing and outstanding balance sync.<h3>Onboarding</h3>15-minute wizard: business profile → agent configuration → WhatsApp channel connection → first AI response live. No technical knowledge required from the client.<h3>Why This Exists</h3>South African SMEs collectively spend R180B+ annually on admin. Most of it happens on WhatsApp, in spreadsheets, and in someone's head. AdminOS is built for a market where the CRM is a contact list, the filing system is a WhatsApp group, and the accountant is the owner's cousin. It does not try to replace that culture. It builds AI infrastructure on top of it.Stack: Next.js 14 · TypeScript · Supabase (PostgreSQL + RLS) · Claude Sonnet (prompt caching) · Upstash Redis · 360dialog WhatsApp API · Resend · PayFast · Xero API · Vercel Edge Functions + Cron · PostHog · Sentry

AdminOS — AI Business OS for SA SMEs

STOKVELOS — AI Stokvel Management PlatformSA's first AI-powered stokvel management platform. R50 billion moves through South African stokvels annually. Most of it is managed via WhatsApp group chats and handwritten ledgers. StokvelOS digitises and protects that wealth.What a Stokvel Actually IsA rotating savings club — typically 10 to 50 members contributing a fixed amount monthly, managed by a voluntary administrator with no formal training and no technical infrastructure. Fraud, miscommunication, and manual errors cost South African stokvels an estimated R3B+ per year. StokvelOS exists to close that gap.Core FeaturesMember management with contribution tracking, payment status, and history per member. AI Health Report — monthly Claude-generated compliance analysis: identifies members behind on contributions, flags unusual patterns, generates a plain-language summary for the administrator. Written to be understood by someone with no financial background.AI Meeting Minutes — administrator enters rough notes from the monthly meeting → Claude generates formatted, professional minutes with action items extracted and assigned. Reduces the administrative burden of running a stokvel from 4 hours to 20 minutes.WhatsApp Reminder System — personalised payment reminders per member, drafted by Claude in the tone and language appropriate to the relationship. Not a generic broadcast. A message that sounds like it came from a person who knows you.Smart Insights — trend detection across contribution history. Flags the early signals of member financial distress before it becomes a missed payment that damages community trust.ArchitectureNext.js 14 App Router · TypeScript · Supabase with RLS (member data fully isolated per stokvel group) · OpenAI GPT-4o-mini for cost efficiency on structured generation tasks · Tailwind CSS · Vercel Analytics and Speed Insights built in.Multi-tenant: each stokvel group is a fully isolated tenant. The administrator of Group A has zero visibility into Group B's data by RLS design.Why OpenAI Here, Not AnthropicStokvelOS uses GPT-4o-mini for its structured generation tasks (minutes, health reports, reminders) because the tasks are templated and deterministic — Haiku-equivalent cost optimisation. AdminOS and VarsityOS use Claude Sonnet for emotionally nuanced responses. The AI choice is made per product based on the nature of the task, not brand loyalty.Stack: Next.js 14 · TypeScript · Supabase · OpenAI GPT-4o-mini · Tailwind CSS · Vercel

StokvelOS — AI Stokvel Management Platform

WATCHSANKOFA — African Creator Streaming PlatformAfrican creator streaming and discovery platform. Named for the Ghanaian Adinkra concept of Sankofa — flying forward while looking back. The premise: African stories, African creators, African audiences deserve a platform that was designed from the ground up for them, not retrofitted from a Western original.The ProblemYouTube takes 45% of ad revenue from African creators. Netflix African content represents under 3% of their catalogue. African filmmakers, musicians, and spoken word artists are distributed on platforms that were not built with them in mind, priced in currencies that hurt them, and governed by content policies that frequently misclassify African cultural content as violating community standards.What WatchSankofa BuildsCreator-first discovery architecture. Upload, profile, and discoverability tools designed for independent African filmmakers and musicians without distribution deals. Community features that connect African creative communities across the continent and diaspora. Content that does not have to be palatable to a Western algorithm to be discoverable.Technical ArchitectureNext.js 14 App Router · TypeScript · Supabase for auth, user data, and creator profiles · Cloudinary for video upload, transformation, and delivery · Zustand for client state · Custom hooks for video player state management · Vercel for deployment.RLS on all user and creator data. Creator upload pipeline: Cloudinary webhook → Supabase update → creator dashboard notification via Realtime.Stack: Next.js 14 · TypeScript · Supabase · Cloudinary · Zustand · Vercel

WatchSankofa — African Creator Streaming Plat

<h3>VARSITYOS — AI Student OS</h3>SA's first AI companion built specifically for South Africa's 11 million university and TVET college students. Not a generic productivity app. A system designed around the specific stressors of the SA student experience: NSFAS payment delays, load shedding during exam season, imposter syndrome in first-generation students, and surviving on a R50/day food budget.<h3>Multi-Agent Architecture</h3>8 independent Claude-powered agents reading from a shared Supabase database. Each agent has a distinct system prompt, a distinct function boundary, and a distinct trigger condition. They do not call each other — they read shared state and act independently.Nova Companion — primary conversational AI. Trained on SA-specific context: NSFAS terminology, university module naming conventions, load shedding vocabulary, South African mental health resources. Surfaces SADAG and Lifeline SA helplines automatically when crisis language is detected in conversation, without the user having to ask.Proactive Insights Agent — runs on dashboard load. Analyses current task completion rate, upcoming exam dates, budget burn rate, and last Nova conversation. Generates one prioritised insight card per session. Never repeating advice.Daily Check-In Agent — triggered at 08:00 SAST. A one-question wellbeing pulse. Response feeds the wellness score graph. Aggregate data across check-ins surfaces burnout risk indicators to the student themselves (never to any third party).Budget Coach Agent — analyses transaction entries against the student's NSFAS allowance. Generates a financial health score 0–100. Produces a plain-language breakdown of spend patterns and a concrete recommendation. Aware of SA-specific spend categories: airtime, taxi fare, grocery staples at Shoprite pricing.Recipe Generator Agent — takes budget amount + dietary preferences + available ingredients → returns a full recipe with estimated cost in ZAR at current Pick n Pay / Shoprite pricing. Built because food insecurity is a real academic risk factor for SA students.Study Planner Agent — generates personalised study plans from module list + exam dates + available hours. Grade calculator included. Exam prep guides generated per module.<h3>Technical Decisions</h3>PWA implementation for offline functionality — load shedding means reliable internet cannot be assumed. Zustand for global state management. React Hook Form + Zod for all input validation. PayFast for ZAR billing (PayPal and Stripe do not support SA bank accounts for most students).Supabase RLS ensures each student's data — including mental health check-ins — is completely isolated. No student's wellness data is accessible to any other user or any admin query.<h3>Pricing Philosophy</h3>Free tier (15 AI messages/month) exists because a student with R0 in their account still deserves access. Paid tiers at R39 and R79/month — deliberately below the cost of one meal at a university cafeteria. Priced for the actual user, not the market average.Stack: Next.js 14 · TypeScript · Supabase · Claude Sonnet + Haiku routing · PayFast · Zustand · React Hook Form + Zod · Vercel · PWA · PostHog · Sentry

VarsityOS — AI Student Super-App

K53 DRILL MASTER — SA Learner's Licence PrepSA's most accessible learner's licence preparation app. 600+ questions across all K53 test categories with spaced repetition logic. First SA learner's test app with isiXhosa language support — built for the 8.2 million Xhosa-speaking South Africans for whom English-only test prep is a barrier to passing, not a knowledge gap.Why This MattersThe South African K53 learner's licence test has a pass rate under 50%. Most failures are not caused by not knowing the rules of the road — they are caused by not understanding the English phrasing of the questions. Language should not be the reason someone cannot drive legally.Technical ApproachDeliberately lightweight: Vite · JavaScript · HTML. No framework overhead. Loads on 2G. Works on entry-level Android devices. Fast, reliable, accessible on any device a learner driver in SA is likely to own.Spaced repetition: questions the user gets wrong appear more frequently. Category-based drilling: road signs, rules of the road, vehicle controls — each practised independently before the full mock test.Planned rebuild in Next.js 14 to match the full Mirembe Muse brand standard, with expanded isiZulu and Sesotho support.

K53 Drill Master — Learner's Licence Prep

<a target="_blank" rel="noopener noreferrer" class="break-all link-break text-gray-500 hover:underline" href="http://CREATIVELYNANDA.CO.ZA">CREATIVELYNANDA.CO.ZA</a> — Personal Brand PlatformPortfolio, blog, Notion template store, poetry section, and AI chatbot built as a unified personal brand platform for Nandawula Kabali-Kagwa, founder of Mirembe Muse. The most iterated repo in the portfolio: 100+ commits, most of which were decisions about what to remove rather than what to add.<h3>Architecture Decisions</h3>Next.js 14 App Router for nested layout architecture — each section (blog, shop, poetry, about) has its own shell without re-rendering the root. Server Components reduce JS payload on a content-heavy site where first-load performance directly affects credibility with international visitors.Multilingual support via Next.js locales. Resend for transactional email with custom React Email templates. PayFast for ZAR template sales. Supabase for contact form submissions, blog content, and Nanda AI chatbot conversation history.The Nanda AI chatbot is Claude-powered with a system prompt trained on Nanda's full professional context — acts as a first-response layer for client enquiries, capable of answering technical questions about the app portfolio, discussing services and pricing, and routing complex enquiries to the contact form.Grain texture overlay implemented as an SVG feTurbulence filter fixed to the viewport — not a repeating background image — to prevent scroll repaints. Design palette: navy, cherry, beige, white with gold as structural accent only.POPIA-compliant across all data collection points. Reg. No. 2026-005658.Stack: Next.js 14 · TypeScript · Supabase · Tailwind CSS · Claude API · Resend · PayFast · Framer Motion · Vercel · Arcjet (per-route, not middleware)