System Architecture · Technical Deep Dive

Engineering Sri Lanka's FirstAI-Native Learning Ecosystem

Client Case Study: ZisuGen | A comprehensive breakdown of multi-agent LLM orchestration, bitemporal RAG search, high-concurrency database isolation, and extreme cloud cost optimization.

Architected by Shehan Chamika•June 2026•15 min read

LTV:CAC Ratio288:1• VC Standard is 3:1

Peak Capacity400 Cont.• Google Cloud Run scaling

Cost Per StudentRs. 0.14• Per Hour serverless budget

Dev Time to MVP2 Months• Supercharged by Antigravity AI

The Paradigm Shift: EdTech Re-Imagined from Scratch

Tuition education in Sri Lanka remains highly traditional, relying on chaotic WhatsApp groups, physical class cards, and fragmented ecosystems. While tutors face a perpetual crisis of content piracy, students struggle with unpersonalized learning, and parents are left completely blind to actual educational progression.

ZisuGen was engineered to dismantle these inefficiencies. It is not a simple "AI wrapper" or standard LMS; it is a fully integrated, multi-tenant ERP, CRM, and AI-Native learning ecosystem. Built on a hybrid architecture combining serverless scalability with relational data integrity, ZisuGen serves as a case study of how custom software can digitize complex real-world workflows.

Architectural Core Stack

Frontend LayerNext.js PWA (Edge)

Backend ComputeNode.js on Cloud Run

Primary DatabaseCloud SQL PostgreSQL

Real-Time SyncFirebase Realtime DB

Vector Searchpgvector extension

Edge MiddlewareCloudflare Workers

AI LLM ModelsGemini 3.1 Pro & Flash

THE PIVOT STORY

From Disaster Relief to Enterprise EdTech: The Sisu Diriya Era

ZisuGen's codebase dates back to December 6, 2025, originally conceived as a rapid-response disaster relief initiative named Sisu Diriya (Student Courage) following the catastrophic Cyclone Ditwah. Children across rural Sri Lanka had lost books, school supplies, and teachers.

To connect donors with children, we engineered a sophisticated Bipartite Geo-Spatial Matching Feed Algorithm. The system parsed unstructured donation/request posts using AI to extract entity metadata (e.g., "Grade 10 Science Textbook in Galle") and dynamically calculated optimal matches:

// Geo-Spatial Matching Radius Expansion Loop

const MATCHING_RADII = [5, 30, 100]; // in kilometers

for (const radius of MATCHING_RADII) {

const match = await locateNearestDonor(studentGeo, bookMetadata, radius);

if (match) { await dispatchRealtimeAlerts(match); break; }

}

Though the disaster relief window closed as schools began physical recovery in early 2026, the core engine—the AI matching, the real-time push notification telemetry, and the deep understanding of student workflows—became the robust foundation for ZisuGen.

The Bipartite Match Formula

Sisu Diriya calculated matches using a combined scoring matrix of geographical distance, subject urgency, and donor validation levels:

MatchScore = W_d*(1 - D/D_max) + W_u*Urgency + W_v*Verification

This ensured that critical requests in remote areas were prioritized and matched with high-trust donors.

CORE INFRASTRUCTURE PILLARS

High-Concurrency Identity, Security & Data Models

Hybrid Authentication

Identity begins at the edge with Firebase Authentication (supporting Google Sign-In and mobile OTP). However, authorization is handled by our Node.js middleware. By parsing the auth token, the backend cross-references our PostgreSQL database via the Firebase Admin SDK to check active student/teacher enrollments, role privileges, and issues a secure, HttpOnly JWT cookie with a 7-Day Sliding Window Protocol to prevent login friction.

Polymorphic Schema

Anticipating loads of 100,000+ concurrent students sitting for national examinations simultaneously, traditional SQL JOINs on the auth path would bottleneck performance. ZisuGen uses a Flattened Polymorphic User Schema (Single Table Inheritance). All roles (Student, Teacher, Parent) occupy a single, heavily indexed User table, yielding O(1) time complexity for initial profile hydration.

Data Store Separation

We optimized cloud billing and latency by splitting storage based on state persistence characteristics. **80%** of the application state (financial transactions, syllabi vector chunks, user profiles) sits in **Cloud SQL PostgreSQL**. The remaining real-time data points (MCQ Game Arena, live attendance triggers, real-time message flags) run on **Firebase Realtime Database** for sub-second synchronization.

RETRIEVAL-AUGMENTED GENERATION

Bitemporal RAG Vector Architecture & Legacy Transcoding

To combat hallucination tendencies in localized education curricula, ZisuGen operates a custom RAG (Retrieval-Augmented Generation) pipeline over a vector database powered by Cloud SQL PostgreSQL and the pgvector extension.

1. Bitemporal Version Control

National syllabi change every few years. To prevent the AI from generating answers based on outdated curriculum standards, our `KnowledgeChunk` vector schema implements a Bitemporal Data Architecture (`validFromYear` and `validUntilYear`). Active syllabi are assigned a temporal infinity value (`9999`). When a new syllabus is ingested, the pipeline automatically mutates the overlapping legacy chunks' expiration timestamps.

2. Legacy Font Transcoding Pipeline

Over 70% of historical Sri Lankan educational materials and past papers are stored in PDFs using non-Unicode, legacy fonts (such as FM Abaya and Bamini). Feeding raw extractions to vector models yields garbage output. We engineered a transcoding middleware that dynamically converts these legacy strings into standardized Unicode, sanitizes line breaks, and routes them to gemini-embedding-2 for vectorization.

Ingestion Engine

RAG Processing Pipeline

Step 1: Legacy Text Extract

"fyksIs kshuhka yd Ndj;hka ms<sn|j" (FM Abaya Legacy Encoding)

Step 2: Transcoding Middleware

"භෞතික විද්‍යාවේ නියමයන් සහ භාවිතයන් පිළිබඳව" (Standard UTF-8 Unicode)

Step 3: Vectorization & Distance Querying

Cosine Distance query optimization:

SELECT chunk_id, content FROM KnowledgeChunk
WHERE country = 'LK' AND valid_until_year = 9999
ORDER BY embedding <=> $1 LIMIT 5;

Note: Pre-filtering on country and temporal validity limits comparison to relevant namespaces, optimizing response latency to < 90ms.

COGNITIVE ORCHESTRATION

Autonomous Agentic Swarms & Multi-Vendor Pipelines

Instead of relying on single LLM invocations, ZisuGen coordinates specialized autonomous swarms. Tasks are dynamically routed depending on cognitive requirements and cost governance thresholds.

01ZisuPedia Authoring Swarm

To generate high-quality educational encyclopedia articles, ZisuPedia employs a 4-Stage Routing Pipeline:

Stage 1: Sanitization (Gemini 2.5 Flash)Extracts keywords, targets key concepts, and structure templates, generating a precise Table of Contents (TOC).
Stage 2: Context Retrieval LayerTriggers localized pgvector vector queries and external real-time scrapes using the Tavily search API.
Stage 3: Deep Drafting (Gemini 3.1 Pro)Utilizes deep reasoning to draft section-by-section based on the TOC, bypassing token limit truncation.
Stage 4: Verification AgentPerforms a final validation pass mapping content against official syllabus blueprints before database publication.

02ZisuDraft Pro & Humanizer Pipeline

For university students, generating draft papers requires highly specific formatting and absolute protection against AI detectors that misidentify student drafts. ZisuDraft implements a unique **Multi-Vendor Model Orchestration**:

1
Initial Drafting (Groq + LLaMA 3.3 70B)Offloaded to LLaMA 3.3 on Groq via high-speed API endpoints to output long-form drafts segment by segment.
2
Asset Generation (Gemini Image Model)Programmatically compiles graphs, mathematical figures, and diagrammatic visuals matching the written concepts.
3
The Humanizer Agent (Gemini 3.1 Pro)Injects the LLaMA text and visual metadata into a final Gemini 3.1 Pro prompt configured with custom journalistic and academic styles. By mixing training datasets, the output achieves an AI-detection score of **1% to 3%** (ZeroGPT baseline) while keeping strict formatting logic.

ERP & FINANCIAL ANTI-FRAUD

The Tuition ERP: Subdomain Routing & AI Slip Validation

Edge Subdomain Routing

To empower educators to build their personal brand, ZisuGen assigns each verified teacher a custom subdomain (e.g., `physics.zisugen.lk`). We engineered edge routing logic using **Cloudflare Workers**. The worker intercepts incoming edge requests, identifies the subdomain namespace, and rewrites the request to the dynamic user route in our central Next.js server without deployment overhead or DNS delays.

AI Slip Verification & Hash-Matching

Manual bank slip verification is slow and open to fraud. ZisuGen uses Gemini Vision models to analyze uploaded slip screenshots for structural indicators (image manipulation, cropped dates, overlaying text, edited values). Once verified, the slip is hashed using a perceptual hashing algorithm. If a student attempts to re-upload the same slip for another month, the database immediately flags the duplicate hash block.

Decentralized Logistics (Tute Control)

High-volume tuition relies heavily on physical notes (Tutes). ZisuGen's logistics engine tracks printing/ink costs, binding expenses, and paper costs to calculate precise P&L margins. When a student collects material, staff scan their digital ID pass via the web application. The system cross-references active monthly fee cards and logs the disbursement in real-time, preventing stock leakage.

AUDITORY EDUCATION & VOICE DSP

Auditory Learning Engines & Noise-Gated DSP

EdTech should accommodate students who spend hours daily commuting on public transport. ZisuGen incorporates the **Auditory Learning Engine** (ZisuAudio), generating bilingual podcasts and audio dramas (dramatizing classical school plays like Sinhabahu with distinct voice profiles).

To enable hands-free voice interactions inside noisy environments, our Live Voice websocket interface (connecting directly to the Gemini Live API model) utilizes a custom Digital Signal Processing (DSP) pipeline.

•Voice Activity Detection (VAD): Configured with hangover frames and exponential smoothing to detect speech patterns.
•Hard Noise Gate (Threshold 6): Aggressively suppresses ambient background noises common in public transit environments.
•Progressive Audio Streaming: Audio generation is chunked. As the first segment is computed, it begins playing immediately, yielding a latency of < 200ms.

Live Voice Connection

Gemini Live Server-Sent WebSockets

DSP Input Frequency

VAD StatusActive Speech

Noise Supp.-24dB (Gate 6)

TELEMETRY & UNIT ECONOMICS

Micro-Psychometric Telemetry & Megascale Scaling

The Micro-Psychometric Data Moat

Standard LMS engines evaluate students based on lagging indicators (final marks, attendance lists). ZisuGen logs how a student learns, capturing micro-interactions per learning session:

•Cognitive Hesitation: Latency metrics measuring exactly how long a student hesitates before picking an MCQ answer option.
•Doubt Mapping (Option Switching): Capturing the patterns of users toggling between specific choices before submitting, indicating conceptual confusion.
•Interactive Friction Points: Timestamp logging tracking when a student pauses video playbacks to trigger the AI Tutor.

Google Cloud Run Auto-Scaling Model

Weekly island-wide mock exams represent a significant concurrency challenge, spiking traffic exponentially at 9:00 AM every Sunday. To mitigate the **Thundering Herd** problem, we modeled scale using:

active_containers = f(U_concurrent / R_max)

Docker containers on Cloud Run scale horizontally instantly, capping compute resources at 400 active instances. Our serverless footprint ensures a minimal baseline cost of just **Rs. 0.14 per student / hour** during exams.

THE ENTIRE ARCHITECTURAL TOOLSET

Next.js 14 / React PWA

Edge routing, dynamic metadata, and offline-first service workers

Node.js (Dockerized)

Microservices containerized and auto-scaled via Google Cloud Run

Cloud SQL PostgreSQL

Active transaction engine with pgvector for localized embedding spaces

Firebase Suite

Firebase RTDB for latency-critical tasks, Firebase Storage for asset storage

Gemini 3.1 Pro & Flash API

Multi-agent authoring, Socratic tutoring, and transcript fetching pipelines

Cloudflare Workers & CDN

Edge rewrite routing for teacher subdomains and network egress optimization

THE OUTCOME

Redefining EdTech in Emerging Markets

ZisuGen is currently live, hosting over 50 real, active daily users in its testing phase. By combining advanced AI-native pipelines with rigorous operational control, the platform has successfully eliminated the geographical and economic barriers of private tutoring.

From a development perspective, ZisuGen showcases the future of engineering. By using the Google Antigravity AI Agent IDE, the entire platform—from database architecture to complex multi-agent swarms—was built, verified, and deployed in a rapid **2-month sprint**.

Build your next system

Looking to engineer your own AI-Native SaaS?

At SoftGen, we transform complex business requirements and raw architectural ideas into secure, scalable, production-grade applications. Contact us today.

Get a Free Consultation

Continue reading more insights from SoftGen

For Business Owners