Tharusha's creative engineering desk with a micromouse robot, monitor, circuit board, playing cards and open sketchbook
PORTFOLIO ONLINECOLOMBO · SRI LANKA
01Projects02Micromouse
Projects
Scroll to explore ↓
FIRMWARE ◆ CONTROL THEORY ◆ POWER CONVERSION ◆ MACHINE INTELLIGENCE ◆ ROBOTICS ◆FIRMWARE ◆ CONTROL THEORY ◆ POWER CONVERSION ◆ MACHINE INTELLIGENCE ◆ ROBOTICS ◆

PROJECT ARCHIVE / 2025–26

Built to understand.
Tested to perform.

Experiments and systems that connect physical hardware, mathematical models and learning algorithms.

10 projects

01Neuromorphic AI
Receptive fields learned by the Synthbrain spiking neural network, forming recognizable handwritten digit patterns
01Project resultSNN-01X / AI / ML
Validated

Synthbrain

A spiking neural network built from scratch in NumPy. It learns MNIST digit structure through spike timing, without backpropagation or labels during training.

90.0% ± 0.7linear-probe accuracy
  • NumPy
  • STDP
  • LIF neurons
  • CUDA
02Control + Robotics
Concept render of the cart-pole test rig with its linear rail, stepper drive, ESP32 controller and upright pendulum
02Concept renderCTRL-02X / ROBOTICS
Building

Cart-pole pendulum

An ESP32 cart-pole platform combining energy-based swing-up, LQR stabilization, encoder feedback, homing and motion safety.

1024 PPRpendulum encoder
  • ESP32
  • LQR
  • C++
  • Mechatronics
03Code Generation
PyGen benchmark chart comparing base and fine-tuned Qwen2.5-Coder-3B results
03Project resultQLR-03X / AI / ML
Validated

PyGen

A reproducible QLoRA pipeline that fine-tunes Qwen2.5-Coder-3B on 18,000 Python instruction-to-code pairs using a single 8 GB laptop GPU.

−73%held-out perplexity
  • Qwen2.5
  • QLoRA
  • PyTorch
  • Gradio
04Computational Neuroscience
Full adult fruit fly brain connectome reconstructed from FlyWire neuron data
04FlyWire / MRC LMBBIO-04X / AI / ML
Validated

Flybrain

A sparse whole-connectome simulation of the adult fruit fly, coupling real neural wiring to learning circuits and an embodied MuJoCo model.

139Kneurons simulated
  • PyTorch
  • Sparse tensors
  • MuJoCo
  • FlyWire
05Scientific ML
An exoplanet illuminated by a distant star against a colorful galactic nebula
05Project visualASTRO-05X / AI / ML
MVP

Exoplanet Hunter AI

A review system that combines NASA catalog classification with light-curve morphology analysis and an explainable candidate queue.

3NASA missions covered
  • scikit-learn
  • FastAPI
  • Astropy
  • MAST
06Reinforcement Learning
Classic Doom logo with blue mechanical details and orange industrial lettering
06Project visualRL-06X / AI / ML
Trained

DOOM Double-DQN

A dueling Double-DQN agent that learns the real Doom engine from raw screen pixels, reward signals and experience replay.

~5kills per life
  • PyTorch
  • ViZDoom
  • Double DQN
  • CNN
07Control Systems
Reaction-wheel pendulum simulation snapshots showing the arm recovering from a disturbance
07Project resultRWIP-07X / ROBOTICS
Simulation complete

Reaction-wheel balancer

A reaction-wheel inverted pendulum model with LQR control, sampled-data validation and active wheel-momentum management.

200 Hzsampled controller
  • MATLAB
  • Simulink
  • LQR
  • ESP32 target
08Connected Embedded
Concept render of the Smart Agriculture system with its LoRa gateway, two sensor nodes, flow sensor and irrigation pump
08Concept renderLORA-08X / EMBEDDED
Prototype

Smart agriculture network

A multi-node 433 MHz LoRa network with acknowledgements, retries, relay automation, local display and cloud telemetry.

3 nodesdistributed system
  • ESP32
  • LoRa
  • Firebase
  • Sensors
09Language Models
Animated attention visualization from the Nano-GPT model showing relationships between tokens
09Project resultLLM-09X / AI / ML
Trained

Nano-GPT

A decoder-only transformer and BPE tokenizer written from scratch, trained locally as a compact instruction model with attention visualizations.

28Mmodel parameters
  • PyTorch
  • Transformers
  • BPE
  • Gradio
10Learning from Scratch
The trained NeuroPong agent sustaining a rally in the Pong environment
10Project resultDQN-10X / AI / ML
Validated

NeuroPong

A NumPy neural network that learns Pong through neuroevolution and DQN, with visualizations of weights, activations and decisions.

8-hitlearned rally
  • NumPy
  • DQN
  • Evolution
  • Pygame
14+public repositories
139Kneurons simulated
90%linear-probe accuracy
4engineering domains

AFTER HOURS / CREATIVE PRACTICE

Engineering is one part of how I make things.

Drawing, 3D modelling and card magic sharpen the same skills I use in technical work: observation, control, patience and iteration.

01 / ART + 3D

Sketchbook and models

Drawings, mechanism studies, visual experiments and 3D-model renders. The gallery is ready for the real work.

02 / CARD MAGIC

Sleight of hand

A private practice in precision, timing and directing attention. Selected routines can be documented here later.

Tharusha PahasaraPROFILE / TP-101
06.9271° N
79.8612° E

ABOUT / SYSTEMS THINKER

Engineering across the boundary between code and physics.

I'm an Electrical & Electronic Engineering student focused on AI engineering, control systems and practical electronics. I care about the full path from equations and sensing to reliable software and measured results.

My projects move between embedded firmware, power conversion, dynamic-system modelling and machine learning. That range helps me see a product as one connected system.

AVAILABLE

AI/ML, embedded and control internships · open-source collaboration · research-adjacent engineering

WORKBENCH / CORE TOOLS

What I build with.

01

Languages

  • Python
  • C / C++
  • MATLAB
  • TypeScript
02

Embedded

  • ESP32
  • STM32
  • LoRa
  • SPI / I²C / UART
03

AI + Data

  • PyTorch
  • NumPy
  • OpenCV
  • scikit-learn
04

Engineering

  • Simulink
  • PCB design
  • Git
  • Linux

LAB LOG / CURRENT DIRECTIONS

What I'm exploring now.

01

Efficient local model adaptation

QLoRA fine-tuning, reproducible evaluation and practical deployment of code models on consumer hardware.

AI / ML
02

Machines that stabilize themselves

Energy-based swing-up, LQR balance, sensor fusion and safe real-time motion control.

CONTROL
03

Intelligence from first principles

Spiking networks, reinforcement learning, compact transformers and biologically grounded simulation.

AI / ML

OPEN CHANNEL / LET'S CONNECT

Have a difficult system to build?

I'm open to internships, research work, technical collaborations and engineering projects.

Send an email GitHub LinkedIn