PSRGAN, a Probabilistic Super-Resolution GAN for Satellite Imagery
4× GAN upscaler for Sentinel-2 tiles with per-pixel uncertainty. 2nd place at the JuFo national final 2026.
Deforestation segmentation in the Amazon with K-Means
Change-detection with K-Means segmentation on PSRGAN super-resolved Sentinel-2 tiles to track forest loss more accurately.
Uncertainty based geoanalytics segmentation for Smart Farming
Downstream segmentation on PSRGAN outputs using per-pixel uncertainty maps for geoanalytics in agriculture.
Spiking CNN on neuromorphic MNIST with population coding
Surrogate-gradient spiking CNN on N-MNIST using population coding (>98% accuracy).
SNN for exoplanet classification using the transit method
Spiking network for low-power detection of exoplanet transits in Kepler/TESS lightcurves.
SNN experiments on MNIST
Basic spiking network experiments on MNIST inspired by the snnTorch tutorials.
Experimenting with Hinton's Forward-Forward algorithm
Reimplementation and experiments with Hinton's Forward-Forward learning algorithm as an alternative to backpropagion.
Spike Encoding algorithms (Rate, Latency, Delta)
Implementation and comparison of rate coding, latency coding, and delta modulation for spike encoding.
Simple neuron simulations
LIF (Lapicque & Alpha) and Izhikevich neuron simulations with snnTorch and Python.
EvidenzLLM: verified LLM responses
RAG system that grounds LLM answers in evidence and refuses when coverage is insufficient via NLI models. BWKI 2025 top-10.
Cenabot: autonomous car to bring snacks
Small autonomous car using a camera, ultrasonic sensor, and wifi-connected RaspberryPi to bring snacks to my laptop.
News sentiment analysis agentic RAG system
Agentic RAG pipeline that retrieves news, scores sentiment, and synthesises a summary using LangChain.
FinBridge: AI financial document analysis
Internship project with fintech startup DocuBridge. Flask webapp to analyze your Excel financial reports. Preprocessing and LLM-based analysis.
CAD LLM-preprocessing webapp
Webapp that converts STEP geometries into relational feature graphs to enable LLM-analysis. Colab with Volkswagen IE.
Lego reinforcement learning robot with no libraries
Distance-based RL neural network for a LEGO SPIKE robot built from scratch with no ML libraries, just pure Math.
CNN for person detection in 3D LiDAR point cloud data
Part of an internship at the German Aerospace Center. CNN trained on a handmade dataset to classify humans in LiDAR point clouds with >98% accuracy.
YOLOv8 Finetuning for wildfire detection
YOLOv8 fine-tuning to detect wildfires and smoke for early combat measures..
Tomato grading binary classification CNN
CNN quality grader for tomatoes: binary classification into fresh/rotten. Agriculture-vision project for quick sorting.
MPRNet for image Deraining
Multi-stage progressive image restoration network (MPRNet) to derain images for autonomous driving.
Research on methods to deal with small datasets
Experimenting and explaining multiple methods to combat data scarcity in a 40-min article. From monte carlo dropout to BNNs and Optimal Transport.
Bayesian Neural Network Test on MNIST
Weight-uncertainty networks trained on MNIST for a quick test to experiment with BNN architectures and methods.
Fish-species classification CNN
Convolutional Neural Network trained to classify 468 species of fish. Transfer Learning with ResNet18 baseline.
Degrees of Separation using Breadth-First Search (BFS)
Find the shortest path between two actors via shared movies using BFS. Inspired the the six degrees of Kevin Bacon.
TicTacToe using Minimax algorithm
Optimal TicTacToe AI using the Minimax algorithm with alpha-beta pruning.
Knights and Knaves logical inference solver
Propositional logic solver for the classic Knights and Knaves puzzles.
Minesweeper AI using propositional logic and knowledge representation
AI agent that plays Minesweeper using propositional logic and a knowledge base.
Google's pageranking algorithm
PageRank via both random-surfer sampling and iterative convergence.
Heredity probability model using Bayesian inference
Bayesian network to compute the probability of genetic traits across Harry Potter family members.
Crossword puzzle generator using constraint satisfaction and backtracking
Generates valid crossword puzzles using arc consistency and backtracking search.
Shopping prediction model using supervised learning
k-NN classifier predicting whether a user will complete an online purchase.
Nim AI using Q-learning
Q-learning agent that learns to play Nim optimally through self-play.
CNN traffic sign classification
Convolutional neural network trained on the GTSRB dataset to classify 43 traffic sign categories.
Natural language sentence parser using Context-Free Grammar (CFG) and NLP
CFG-based parser that identifies noun-phrase chunks in English sentences.
Testing simple attention mechanisms in a question answering system
TF-IDF + attention-based QA system that extracts answers from a document corpus.