Karl Sedem
Haligah

📋 Problem

A wearable device produces a continuous stream of integer heart rate readings. Within this stream, certain subsequences correspond to dangerous arrhythmia episodes that require immediate medical attention. The challenge is to automatically locate these episodes, which are fixed-length intervals of 50 measurements, within a much longer signal containing both normal and abnormal rhythms.

⚙️ Approach

• Trained two separate first-order Markov models — one for normal heart rhythm (N) and one for arrhythmia episodes (A) — from labeled training sequences
• Discretized raw heart rate values into 8 state bins to reduce noise and enable reliable transition probability estimation
• Applied Laplace smoothing to avoid zero-probability transitions and used log-space arithmetic to prevent numerical underflow
• Classified using a sliding window (length 50) over the evaluation signal, computing a log-likelihood ratio test at each position with prior probabilities P(A)=0.05, P(N)=0.95
• Selected final detections via greedy non-overlapping interval selection, ranked by classification confidence

📊 Results

Evaluated on three test sequences. A detected interval counts as correct if it overlaps at least 80% with a true arrhythmia interval.

The consistently high recall shows the system successfully identifies the majority of true arrhythmia episodes — critical for a medical monitoring context where missing a dangerous event is far worse than a false alarm.

🛠️ Tech Stack

📋 Problem

Given a bus network where multiple lines operate on fixed intervals, a passenger arrives at a start stop at a specific time and needs to reach a destination stop as quickly as possible. The total cost includes the initial waiting time, all travel times between stops, and any transfer waiting times when switching lines. The planner must find the route with the lowest total cost while correctly handling time-dependent departures.

⚙️ Approach

• Modeled the search space with states defined as (stop, time) pairs, since the same stop at different times produces different waiting costs and available connections
• Implemented A* search with a priority queue ordered by f(n) = g(n) + h(n) and a best-g map to avoid re-expanding suboptimal states
• For small test networks, used provided admissible heuristic values; for the large real-world graph, computed heuristics via Haversine distance (straight-line geographic distance)
• Handled periodic bus schedules by calculating the next valid departure from any stop at any given time based on the line's interval and first departure
• Explored a slightly inadmissible weighted heuristic (factor 1.03) to reduce search effort while keeping solutions within 0.01% of optimal

📊 Results

All seven synthetic test cases returned the correct optimal costs, including correctly identifying an unsolvable case. On the large Lübeck bus network (100 test cases):

The weighted heuristic maintained optimal solution quality while slightly reducing search overhead, confirming that a carefully tuned inadmissible heuristic can improve efficiency without meaningful loss in solution quality.

🛠️ Tech Stack

📋 Problem

Tracking personal expenses is tedious when done manually, and most people lose track of where their money goes. Physical receipts pile up and are rarely reviewed. The goal was to build a tool that makes expense entry as frictionless as possible, whether by typing, pasting receipt text, or simply photographing a receipt, and then automatically extract the relevant data and present it in a clear, interactive dashboard.

⚙️ Approach

• Built a modular backend with clean separation: receipt parsing, expense management, data loading, and visualization each in their own module
• Implemented a regex-based receipt parser that extracts total amount (prioritizing SUMME/GESAMT lines), date (supporting multiple formats including OCR artifacts), merchant name, and expense category
• Integrated Tesseract OCR with OpenCV preprocessing (grayscale conversion, Gaussian blur, Otsu thresholding) to extract text from receipt photos
• Created an interactive Streamlit dashboard with KPI metrics (income, expenses, savings), category donut charts, weekly spending bar charts, and cashflow timelines
• Supported full CRUD operations on expense and income records directly from the dashboard, with CSV-based persistent storage

📊 Features

• Three input modes: manual form entry, text paste with auto-parsing, and camera/photo OCR scanning
• Smart category detection based on merchant keywords (supermarkets, transport, drugstores, restaurants)
• Time-based filtering (week, month, year, all-time) across all visualizations
• Editable data tables for correcting or deleting entries directly in the browser
• CLI version available for quick terminal-based receipt logging

🛠️ Tech Stack

📋 Problem

Processing audio content manually is time-consuming, whether it's transcribing a lecture, translating an interview, or extracting key points from a meeting recording. The goal was to build a full-stack platform where users can upload any audio file and instantly get a transcription, a translation into their target language, and a concise summary, with user accounts so results are saved and accessible later.

⚙️ Approach

• Designed a three-tier cloud-native architecture: Streamlit frontend, FastAPI REST backend, and PostgreSQL database, each running in its own Docker container
• Integrated Whisper (whisper-3-large) for speech-to-text transcription and LLaMA (llama-33-70b) for both translation and summarization via an OpenAI-compatible API
• Implemented user authentication with fastapi-users, allowing registration, login, session history, and account deletion
• Deployed the full stack to Kubernetes using Deployments, StatefulSets, Services, Ingress with TLS, PersistentVolumeClaims, and Secrets
• Built a GitLab CI/CD pipeline with four stages (Prepare, Build, Test, Deploy) that automatically builds Docker images, runs tests, and deploys to the cluster on every commit

📊 Architecture

The user uploads audio through the Streamlit UI, which sends it to the FastAPI backend. The backend calls the Whisper model for transcription, then passes the text through LLaMA for translation and summarization. Results are stored in PostgreSQL per user session and can be retrieved or exported as PDF at any time. The entire stack is orchestrated by Kubernetes with automated deployments through GitLab CI/CD.

🛠️ Tech Stack

📋 Problem

Pet owners often struggle to find reliable, short-notice care for their animals, while willing caregivers have no easy way to offer their services. The goal was to build a platform that brings both sides together with trust-building features like verified profiles, star ratings, direct messaging, and calendar-based availability, so arranging pet care becomes as simple as a few clicks.

⚙️ Approach

• Designed and led a team of 4 using Scrum with sprint-based development, backlog management, and regular retrospectives
• Built the backend as a Spring Boot REST API with layered MVC architecture: controllers, services, and JPA repositories with PostgreSQL
• Implemented user authentication and role-based access (pet owner vs. caregiver) using Spring Security
• Developed real-time chat using WebSocket with STOMP over SockJS, including push notifications for new messages
• Created a review system with 1-5 star ratings, text comments, and average score display on caregiver profiles
• Built caregiver search with location/postal code filtering and calendar-based availability views using FullCalendar
• Rendered dynamic frontend pages with Thymeleaf, HTML, CSS, and JavaScript

📊 Key Features

• User profiles with pet photos (up to 4 per pet), bios, and service descriptions
• Real-time WebSocket chat with conversation history and notification badges
• Star rating and review system with per-caregiver averages
• Emergency button for instant caregiver/vet alerts with location sharing
• Availability calendar for caregivers to manage and display their schedule
• Contact form for platform support

📸 Images

‹ ›

🛠️ Tech Stack

📋 Concept

The game combines arcade shooting with environmental puzzle elements. The player remains stationary and uses a flashlight beam to hunt glowing ghost enemies orbiting the room. Some ghosts are hidden inside physics objects such as bottles, lamps, and statues. When a hidden ghost is revealed, it triggers physics interactions that can uncover additional enemies. Invisible ghosts can only be detected by observing nearby light sources that flicker in response to ghost proximity.

⚙️ Technical Highlights

• A custom emissive ghost shader material was developed using material functions, allowing base colors to be saturated or overridden while preserving secondary features like eyes and mouths
• An "orbit" coordinate system was implemented so revealed ghosts move on a cylindrical hull around the player, simplifying movement calculation and eliminating the need for 3D pathfinding or navigation meshes
• A state machine was built in Blueprints to control ghost behavior types (Orbiter, Lurker), with each type reacting differently to the flashlight through freezing, drifting away, or dashing out of the beam
• A modular light interaction system was created using a reusable Actor Component (BPC_LightFlicker) and a Blueprint Interface (BPI_LightInterference), allowing ghosts to dynamically affect any nearby light source based on proximity without rigid class dependencies
• Performance was optimized by switching to forward rendering with dynamic/stationary lighting and using gobo light functions to fake window shadows on the floor, achieving a stable 60 FPS with 30+ active ghosts
• Custom sound design was produced from human voice samples processed with pitch shifts, reverb, and reverse reverb in Tenacity

🎨 Contribution: Blueprint Systems, Level Design & Sound Design

Blueprints for interactive level elements were designed and implemented, supporting item functionality and gameplay interactions within the level. Parts of the level design and environment dressing were developed, contributing to the stylized haunted room atmosphere and overall gameplay readability. Portions of the sound design were produced, including processed ghost voice effects and environmental audio to enhance atmosphere and player feedback.

📸 Images

‹ ›

🛠️ Tech Stack