Hi, I'm Savindu Suranjith

Full-Stack & AI Engineer

Building scalable, user-centric solutions with machine learning, modern web technologies, and autonomous LLM agents.

View Projects Contact Me
Full Stack LLM Agents Computer Vision
Neural Mesh Active & Interactive
Interactive 3D Move & Click to Pulse
React Native Core
Frontend & Mobile Engineering
Hover tech stacks or drag 3D mesh to explore

About Me

Savindu Suranjith Thennakoon

Savindu Thennakoon

CSE Undergrad @ Uni of Moratuwa

Sri Lanka 3rd Year

I am a driven Computer Science and Engineering undergraduate at the University of Moratuwa, specializing in full-stack development, mobile engineering, and applied AI.

My passion lies in bridging the gap between cutting-edge AI research and practical applications. I have extensive experience building cross-platform mobile and web apps integrated with machine learning models and computer vision pipelines.

I am always looking for opportunities to apply robust problem-solving skills to innovative technical projects.

Full-Stack

MERN, React Native

Applied AI

LLM Agents, Computer Vision

Technical Stack

AI & Neural Systems

Applied AI research, on-device vision, and autonomous agent orchestration.

LangGraph (Agents)
ChromaDB Vector Store
MediaPipe Pose
TensorFlow.js
RAG Architectures
Google GenAI / LLMs

Web & Mobile Apps

Responsive UI/UX, state management, and mobile cross-compilation.

React Native
React.js
Node.js / Express
Capacitor
RESTful APIs
Tailwind & CSS3

Databases & DevOps

Resilient data stores, containerization, and API verification pipelines.

PostgreSQL
MongoDB
Docker Container
Git & GitHub
Postman
MySQL

Languages & System Architecture

Core programming languages, algorithmic problem solving, and microservices design.

Python
C++
JavaScript (ES6+)
TypeScript
Data Structures & Algorithms
Microservices Architecture

Featured Projects

CivicGuard (Disaster Response & Emergency Intelligence Platform)

  • Platform Architecture: Built a nationwide disaster response platform utilizing an event-driven microservices architecture (Node.js, FastAPI, Kong Gateway, React 18) and PostGIS to automate incident triage and coordinate 250 tactical field squads across 25 administrative districts.
  • 5-Signal Hybrid Verification: Engineered an automated triage engine under a < 1.5s SLA coupling Gemini 3.5 Flash-Lite multimodal reasoning with YOLOv8 spatial telemetry, PostGIS geospatial clustering, and live hydrological sensor feeds.
  • Concurrency & Systems: Prevented race-condition over-allocation across evacuation shelters during sudden population surges by designing an atomic bed reservation engine using PostgreSQL transactional row-level locking.
TypeScript Node.js Express Python FastAPI React 18 PostgreSQL PostGIS Supabase Google Gemini 3.5 Flash-Lite YOLOv8 Kong API Gateway Docker Compose Socket.IO
View Repository

Smart Attendance and Classroom Access Platform

  • Platform Architecture: Built an enterprise-grade academic attendance platform with a microservices architecture using FastAPI, Next.js, React Native, PostgreSQL, and RabbitMQ.
  • Edge AI & Anti-Spoofing: Designed an on-device biometric verification pipeline leveraging MiniFASNet to detect presentation attacks in real time, computing 192D embeddings locally to preserve privacy.
  • Fraud Prevention & Scalability: Combined biometric checks with GPS geofencing and randomized verification windows; offloaded vector matching to distributed RabbitMQ message queues for high-volume check-ins.
React Native Next.js / React FastAPI PostgreSQL RabbitMQ Supabase MiniFASNet (Edge AI) GPS Geofencing Tailwind CSS
View Organization

FitLock (AI App-Locker)

  • Wellbeing App Lock: Developed a cross-platform digital wellbeing application that locks target apps behind physical exercise requirements (squats, push-ups, jumping jacks).
  • On-Device Computer Vision: Integrated MediaPipe Pose via HTML5 Canvas for client-side, zero-latency repetition counting and joint angle calculation without streaming or storing user video.
  • Monetization & Privacy: Integrated Google AdMob alongside Google User Messaging Platform (UMP) SDK to handle monetization with full GDPR consent compliance. Live on Google Play.
React Native React Vite Capacitor MediaPipe Pose Tailwind CSS Google AdMob & UMP
Google Play

HackerRank Orchestrate AI Agent (Global Rank #81 / 1,400+)

  • Deterministic State Machine: Built a multi-node LangGraph workflow orchestrating classification, safety checks, policy routing, and ticket resolution across HackerRank, Claude, and Visa domains.
  • Local RAG Pipeline: Indexed multi-domain support corpora into an embedded ChromaDB using HuggingFace all-MiniLM-L6-v2 embeddings, ensuring grounded, hallucination-free retrieval without external web dependencies.
  • Type-Safe Routing & Resilience: Enforced strict Pydantic structured output models with asynchronous processing via asyncio.Semaphore and Tenacity exponential retry policies.
Python LangGraph Google GenAI ChromaDB HuggingFace Pydantic Asyncio
View Repository

PrajaNavigator (AI Public Administration Platform)

  • Platform Overview: Engineered a full-stack, multi-agent public administration platform to streamline citizen services via multimodal OCR verification, pre-counter readiness scoring, and automated form generation.
  • Hierarchical RAG Engine: Designed a 3-layer Hierarchical RAG system incorporating automated conflict resolution and multi-agent workflow orchestration across statutory acts, ministry circulars, and Grama Niladhari guidelines.
Python FastAPI Hierarchical RAG OCR & Vision Multi-Agent Systems React 18 Tailwind CSS Docker Compose
View Repository

Extra Curricular Activities

Education

BSc (Hons) in Computer Science & Engineering

University of Moratuwa

2024 - Present (3rd Year)

GCE Advanced Level

Dharmapala College Bandarawela

2022 | 3 As

Leadership & Societies

IEEE RAS

Marketing Pillar Member

MoraSpirit

Marketing Pillar Member

IET On Campus

Operations Pillar Member

Taekwondo Team Member

University of Moratuwa (2025 - Present)

Get In Touch

I'm currently seeking a software engineering internship to apply my problem-solving skills to innovative projects. Feel free to reach out!

Download Resume
SYSTEM ARCHITECTURE DEEP-DIVE
View Repository

Microservices Event-Driven Topology

Full-Stack System Design

CivicGuard decouples emergency intake, AI verification, tactical squad dispatch, and shelter bed allocation across high-throughput containerized microservices. The Kong API Gateway manages routing, token validation, and rate-limiting, while FastAPI handles ingestion telemetry.

Architecture Diagram
CivicGuard System Architecture Diagram
Figure 1.1: Citizen Intake → Kong Gateway → 5-Signal Verification → Event Broker → Tactical Squad Dispatch & Atomic Shelter Lock.
Ingestion & Gateway Layer

Incoming citizen hazard reports and multimedia feeds enter through Kong API Gateway, enforcing strict rate-limiting and TLS termination before reaching the asynchronous FastAPI ingestion services.

Spatial Clustering & Broker

PostGIS spatial clustering algorithms aggregate concurrent reports originating from identical geographic radii, preventing dispatch duplication and broadcasting verified incidents via Socket.IO.

Tactical Command Center: 4 Workspaces

Mission Control Interface

The CivicGuard command dashboard coordinates 250 field squads across 25 administrative districts, unifying situational awareness into 4 purpose-built real-time workspaces.

Command Center Workspaces
CivicGuard 4 Workspaces Command Dashboard
Figure 1.2: Multi-district command dashboard showcasing live hazard telemetry, tactical squad dispatch, and shelter capacity tracking.
01
Citizen Hazard Telemetry Feed

Real-time intake stream displaying citizen hazard submissions, geo-coordinates, and uploaded imagery.

02
Multimodal Incident Triage

Live AI verification workspace highlighting Gemini 3.5 Flash-Lite scene classification and YOLOv8 hazard bounding boxes.

03
250 Tactical Squad Dispatch

Interactive district operational map tracking squad assignments, status, and ETA across 25 administrative zones.

04
Shelter Capacity & Evacuation

Live bed reservation dashboard backed by PostgreSQL row-level locks to avoid capacity collisions during surges.

5-Signal Hybrid Verification Engine (< 1.5s SLA)

< 1.5s Inference SLA

To combat false alarms, hallucinations, and duplicate citizen reports during crises, CivicGuard couples multimodal neural reasoning with spatial telemetry and hydrological sensor feeds.

Intelligence Engine Architecture
CivicGuard 5-Signal Intelligence Engine Diagram
Figure 1.3: 5-Signal triage evaluation pipeline integrating Gemini 3.5 Flash-Lite, YOLOv8 telemetry, and PostGIS deduplication.
01
Multimodal Triage Gemini 3.5 Flash-Lite scene reasoning & damage severity estimation
→
02
Spatial Telemetry YOLOv8 object bounding boxes (floods, collapses, fires)
→
03
Geospatial Clustering PostGIS clustering & deduplication of concurrent calls
→
04
Hydrological Feeds Live water stream sensor telemetry validation
→
05
Atomic Dispatch Row-locked shelter bed allocation & tactical squad deploy

Systems, Concurrency & Operational Metrics

Database Reliability
250 Tactical Squads Coordinated
25 Administrative Districts
< 1.5s Inference SLA
Zero Conflict Row-Level Atomic PostgreSQL Locking
Atomic Bed Reservation & Concurrency Management

During sudden mass evacuation surges, multiple relief officers and field squads simultaneously allocate beds across designated emergency shelters. To eliminate race-condition over-allocations and double-bookings:

  • Row-Level Locking: Uses PostgreSQL transactional row-level locks (SELECT ... FOR UPDATE) on specific shelter room inventories.
  • Atomic Transactions: Ensures reservation decrements, guest registration, and confirmation emit atomically within an isolated transaction boundary.
  • Zero Double-Allocation Guarantee: Prevents overcrowding at shelters and guarantees that emergency capacity figures displayed in the Tactical Command Center reflect ground reality.
Visual architecture diagrams and screenshots in preparation

Edge AI Biometric Verification & Anti-Spoofing

Privacy-by-Design

Eliminates proxy attendance and presentation attacks in university lecture halls by coupling on-device MiniFASNet facial landmark tracking with randomized GPS geofencing.

On-Device Vector Extraction

Computes 192-dimensional vector embeddings directly inside the client device memory using MiniFASNet. Raw facial imagery never leaves the device or travels across network wires.

Asynchronous Queue Architecture

Check-in spikes from 500+ simultaneous students are buffered through distributed RabbitMQ message queues to worker pools, preventing API bottlenecks.

Visual architecture diagrams and screenshots in preparation

On-Device MediaPipe Kinematics & Android Interceptor

Live on Google Play

FitLock tackles phone addiction by locking distracting apps behind verified physical exercises (squats, push-ups, jumping jacks), executing computer vision on-device without video transmission.

33 3D Skeletal Landmark Tracking

Calculates angular biomechanics between hip, knee, and ankle joints in real time via HTML5 Canvas, distinguishing partial reps from full-range exercises.

Android Lifecycle Interception

Native Android accessibility background services intercept target application launches and display the exercise challenge overlay.

Visual architecture diagrams and screenshots in preparation

Multi-Node LangGraph Deterministic State Machine

Global Rank #81 / 1,400+

An autonomous multi-node agentic workflow built with LangGraph, ChromaDB, and Google GenAI to resolve customer support tickets across HackerRank, Claude, and Visa domains without external web dependencies.

Embedded ChromaDB RAG

Indexes multi-domain knowledge bases using HuggingFace all-MiniLM-L6-v2 embeddings for grounded, zero-hallucination document synthesis.

Type-Safe Pydantic Routing

Enforces strict output schemas with Pydantic (.with_structured_output), asyncio.Semaphore concurrency, and Tenacity exponential retry.

Visual architecture diagrams and screenshots in preparation

3-Layer Hierarchical RAG & Citizen Flow

Modernizes public administration by guiding citizens through a 5-step automated documentation workflow powered by a 3-layer Hierarchical RAG engine.

01
Identify ServiceNLP intent classifier
→
02
Dynamic MCQsContextual questionnaire
→
03
Create FormsSchema generator
→
04
Document AutofillMultimodal OCR
→
05
Visit RoadmapPre-counter checklist
High-resolution architecture diagram