Shipped Systems.
Verifiable Outcomes.

1. The Challenge

Immigration eligibility rules live scattered across government sites and PDFs, so most applicants only discover them after they've already committed to a country. Consultancies lose advisor time on unqualified leads with no self-serve way to filter early interest, and a catalog of hard-coded rules goes stale the moment a country updates its points table.

2. The Solution

We built a guided profile wizard - fillable by hand or auto-filled from an uploaded CV - that runs against deterministic, country-specific scoring engines (Canada Express Entry and Skilled Worker, Australia, New Zealand, USA) rather than letting an LLM freehand eligibility decisions. Applicants get a free ranked snapshot across five programs, then unlock a paid deep-dive report via Stripe. Behind the scenes, an ingestion worker proposes rule updates from official sources, and a human reviews and approves every diff before it goes live.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

Direct Home lets Hong Kong property owners list and sell or rent without an agent - which means the platform has to do an agent's job. Listings needed a real draft-to-approval workflow instead of a form that saves straight live, trust had to be built in with verified email, phone, and HKID before a deal could move forward, negotiation needed to live inside one real-time thread instead of scattered emails and calls, and legal paperwork like letters of intent and tenancy agreements couldn't stay manual.

2. The Solution

We built the verified-first backend and web app end to end. Email and phone OTP, Google OAuth, and HKID/identity/KYC review gate most of the API, so by the time two parties are negotiating, both are already verified. A dedicated Socket.IO chat namespace keeps offers, counters, viewings, files, and support in one running thread, Stripe Checkout runs behind a backend webhook rather than client-side Stripe.js, and letters of intent and tenancy agreements generate straight from platform data.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

SmartGuy runs a faith-and-wisdom content platform for consumers built around self-assessments, alongside a US business network where professionals pay for listing slots and earn referral commissions - two very different audiences on one system. Each needed distinct signup, verification, and dashboard paths; business listings needed real recurring monetization rather than a flat one-time fee; personalization needed to follow from real assessment data; and email at scale needed proper queuing and unsubscribe handling.

2. The Solution

We built the backend and web app that keep both audiences on one shared data model without getting in each other's way. Self, friends-family, and idol-match assessments feed a recommendation layer that surfaces trait-matched videos, tasks, movies, and readings. Authorize.Net ARB handles recurring monthly and annual listing-slot subscriptions, including trial and referral-discount packages, and inspirational and campaign emails run through cron-driven queues with proper unsubscribe tokens, kept separate from transactional essentials like verification and password reset.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

Cambridge O & A Level exam boards publish past papers, mark schemes, and examiner reports as disconnected static PDFs, so students rarely connect the exact insight that explains lost marks. Private tutoring can fill the gap but is costly and inconsistent, and students often know the material yet still lose marks purely on technique.

2. The Solution

We built a mobile AI tutor across three connected study loops, all grounded in real exam material: students photograph or type a past-paper question and get a structured, examiner-style answer with a marks breakdown and source citations; their own written or photographed answer gets marked the same way, awarded vs. total with strengths and missed points; and a live quiz checker builds topical or mixed sets on demand and marks every answer as the student goes. If the indexed source material can't support an answer, the system declines rather than inventing one.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

Professional services intake is highly manual, relying on unorganized client briefs, scanned files, and emails. Sorting client requirements, mapping unstructured text formats, and manually entering data into CRMs takes hours, delaying response times and leading to transcription errors.

2. The Solution

We engineered an AI ingestion and document processing pipeline. The system automatically extracts key client details, service queries, and requirements from incoming briefs, structuring them into unified JSON schemas and routing them straight to CRM databases.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

Naturals Club & Resort ran membership, bookings, invoices, and helpdesk tickets entirely through Odoo, its ERP - workable for accounting but a self-service-free experience for everyone else. Members had to call the front desk to book a court or add a guest, staff navigated screens built for accountants rather than a calendar or a gate, and there was no fast, contactless way to bring a guest onto the property.

2. The Solution

We paired a members-only mobile app with a permissioned staff dashboard, keeping Odoo as the system of record behind both. Members book sports, spa, and event slots, generate an 8-hour QR guest pass, pay via One Link with proof upload, and track a shared family wallet - all from their phone. Staff see the same bookings on a calendar, manage resources and pricing, and review payments and feedback, each scoped to their own role.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

AISIS needed a Student Information System MVP a real school district could pilot, where a teacher, a counsellor, and a SpEd coordinator each see only what their role needs. Compliance tracking for SpEd/504 cases needed to be a real workflow rather than a spreadsheet, attendance changes needed a defensible audit trail, and the platform had to isolate each district's data cleanly from day one.

2. The Solution

We built the multi-tenant platform underneath the pilot: a shared Zod contracts package defines the RBAC matrix once so permissions can't drift between frontend and backend, every request resolves its tenant from the subdomain and cross-checks it against the JWT's tenant claim, and four AI features (natural-language search, risk analysis, narrative reports, compliance review) run through an optional Anthropic integration with every output tied back to source-field citations.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

PodcastOne runs one of the largest podcast networks in the industry - hundreds of shows and a constant flow of new episodes. Its legacy CMS couldn't structure shows, episodes, hosts, and advertisers the way the network actually works, episode ingestion from RSS feeds was manual, and the public site needed to stay fast and SEO-friendly as the catalog kept growing.

2. The Solution

We rebuilt the CMS and the public-facing web app as a matched pair. A custom WordPress plugin with dedicated post types and a Field API replaced a generic plugin-based approach, automated WP-CLI ingestion pulls the last 365 days of episodes without overwriting editors' existing choices, and a headless split lets the CMS and the Next.js site deploy and scale independently.

3. System Architecture

4. Business Outcome

Technology Stack

1. The Challenge

Generic quiz apps can't render the interactive item types a real Microsoft certification exam uses - hotspot, drag-and-drop, case studies, active screens, simulated labs - and most question banks are leaked or brain-dumped, with no per-domain diagnosis and no feel for real exam pacing under time pressure.

2. The Solution

We built a Next.js learner and admin product on top of a FastAPI backend. Every session samples from published, human-reviewed questions only, grades each of the nine interactive item types against a stored correct payload, and computes a scaled 100–1000 estimate from the pack's own min/max and passing score. New certification packs ship as data files with no engine changes, and AI-drafted questions stay in a pending queue, scored for quality, until an admin accepts them.

3. System Architecture

4. Business Outcome

Technology Stack