GPA: 27/30 · Thesis: Safety control for LLMs with probabilistic circuits
Featured Links
- Depth-Only Image Classification for Privacy
- Speaker-Aware Meeting Transcription
- Warm-Start Graph Cuts for Video Segmentation
Summary
I work on large language models: getting them to do useful work, keeping them under control, and running them on infrastructure you own.
Right now
- My MSc thesis is on safety control for LLMs using probabilistic circuits.
What I’ve done
- SPRITZ research group: used LLMs to find security issues in Internet protocol specifications.
- Var Group: built an LLM debugging system on a locally deployed model that cut issue analysis to under 12 minutes.
- Before my master’s: software engineer in Addis Ababa, building web platforms and backend services.
My master’s covered deep learning, probabilistic modelling and NLP. I’ve applied that to speech, vision and robotics projects.
Interests
- AI: LLMs, AI safety and alignment, interpretability, NLP, deep learning, computer vision
- Systems: LLM systems (RAG, agents, retrieval pipelines), self-hosted deployment and inference optimisation, ML engineering and MLOps
I graduate in April 2027 and I’m looking for AI/ML engineering and research roles. Get in touch
Skills
- Languages: Python, C++, Java, JavaScript/TypeScript, SQL, C#
- AI / ML: PyTorch, TensorFlow, scikit-learn, Hugging Face, OpenCV, LLMs, NumPy, pandas, LangGraph, LangChain, RAG, Ollama, Sentence-Transformers
- Backend & Databases: Node.js, Express.js, NestJS, Flask, FastAPI, GraphQL, MongoDB, MySQL
- Frontend: React, Next.js, Angular, Vue.js, Redux, Tailwind CSS, MUI, Jest, Cypress
- DevOps & Cloud: Docker, Jenkins, GitHub Actions, GCP, Linux, Git
- Robotics: ROS 2, MoveIt 2, Nav2, Gazebo, Behavior Trees
- Spoken Languages: 🇬🇧 English (fluent), 🇪🇹 Amharic (native), 🇮🇹 Italian (A1)
Education
GPA: 3.79/4.0 (Very Great Distinction)
Competitive Programming and Problem‑SolvingLeetCode
Feb 2023 – Oct 2023
- Completed an intensive programme preparing a selected cohort for engineering roles at top tech companies, combining daily DS&A practice, timed contests and team software projects
- Solved 620+ LeetCode and 100+ Codeforces problems
Professional Experience
Research Assistant
Nov 2025 – Jul 2026
- Designed an LLM pipeline that finds bugs and security vulnerabilities in RFCs (the documents that define Internet protocols)
- Benchmarked retrieval-based and reasoning-based LLM approaches on long technical documents that change over time
- Built a debugging assistant that traces bugs to their root causes and suggests code fixes, running on a self-hosted GPT-OSS-120B model with LangGraph
- Designed a search pipeline that combines keyword search, semantic search and the repository’s code structure to find the right context for each issue
- Reduced issue analysis from manual investigation to under 12 minutes per issue, and replaced paid external LLM APIs with self-hosted inference
Software Engineer
Jun 2023 – Oct 2024
- Led frontend development for hakimhub.app, a healthcare platform connecting patients, hospitals and doctors
- Built sign-in, user profiles and a real-time chatbot that helps patients describe symptoms and find the right doctor, using Next.js, TypeScript, Redux Toolkit and Tailwind CSS
- Automated end-to-end and unit testing with Cypress and Jest, reducing manual QA
- Built a system that automatically detects ads in long audio and video recordings, replacing manual review
- Used audio fingerprinting to match known ads, reducing several minutes of manual work per recording to about a minute
- Analysed gene-expression data to find patterns linked to ageing and longevity, as part of a bioinformatics research team, using pandas, NumPy and Matplotlib
- Built and trained machine learning models on this data with TensorFlow and scikit-learn
- Built and tested an email service with Node.js and SendGrid that sends PDF attachments generated with Puppeteer and Handlebars, packaged with Docker
- Reduced CPU usage by over 30% by replacing a legacy Spring Boot service with Node.js
Projects
Automatic Extraction of Anatomical Ear Landmarks
Aug 2026 – Sep 2026
Huawei Munich Tech Arena 2026·Team project
- Built a system that finds 85 precise points on the ear from a 3D head scan, used to personalise spatial audio without lengthy acoustic measurements
- Located 15 key points first with an ensemble of three deep learning models, then traced the remaining 70 along the ear’s natural curves
- Reached 1.25 mm average error on unseen scans, and showed the 15 key points were the bottleneck: with them perfect, error drops to 0.63 mm
- Ranked 12th of 38 teams as Master’s students, in a challenge aimed at PhD-level researchers
Jul 2026 – Present
Master's Course Project·Speech Processing
- Building, from scratch, a system that turns a meeting recording into a transcript showing who said what
- Chose each part (speech detection, speaker recognition, clustering, transcription) by running experiments rather than using a ready-made tool
- Reached 28.4% speaker-labelling error and about 26% word error on meetings with speakers never seen in training
- Found most errors come from people talking over each other rather than voices being mixed up, which points to the next improvement
Jan 2026
Master's Course Project·Intelligent Robotics
- Built a system where a simulated UR5 robot arm finds two cubes from the tags on top of them, swaps them between two tables, and reports their colours
- Broke the task into separate steps (pick, place, return home), with a behaviour tree deciding what runs next
- Reached a 75% task completion rate, with failed picks and places retried automatically
Dec 2025 – Jan 2026
Master's Course Project·Intelligent Robotics
- Built a system where a simulated mobile robot finds its destination from two AprilTag markers, drives there on its own, and spots the tables around it
- Used Nav2 for driving and the robot’s laser scanner to detect the tables, with automatic retries if navigation fails
Mar 2025 – Jun 2025
Master's Course Project·Vision and Cognitive Systems
- Tested whether standard image-recognition networks can identify objects from depth images alone, which show an object’s shape but no faces, colours or textures
- Adapted four pretrained networks (AlexNet, VGG19, ResNet50, Inception-v3) to depth input, on both real depth scans and depth estimated from ImageNet photos
- Reached 92% accuracy on real depth scans of 51 object types, and 49% on 1,000 ImageNet classes, where full fine-tuning nearly tripled accuracy
Nov 2025
Master's Course Project·Advanced Algorithms
- Reproduced Predictive Flows for Faster Ford-Fulkerson (Wang et al., 2023), which speeds up video segmentation by reusing the previous frame’s solution instead of solving every frame from scratch
- Raised the speedup from 24% to 42% by fixing a bug in the authors’ code that read the video frames out of order
Honours and Awards
Awarded by the Italian Ministry of Foreign Affairs to fund my MSc in Computer Science (AI) at the University of Padova, for academic excellence and leadership potential.
Graduated with Very Great Distinction, achieving a GPA of 3.79/4.0.
Volunteering
STEM Tutor
Local community, Addis Ababa
Tutored underprivileged students in mathematics and programming.
Book Cover Designer
Designed the cover for መጽሐፈ ሐዊ ቅጽ 3 (Mets’hafe Hawi, Vol. 3), an Amharic religious book transcribed from ancient Ge’ez manuscripts.