Haben Eyasu

AI & Data Engineer

I build data-driven AI systems, from production pipelines and machine learning models to privacy-first, on-device LLM applications.

AI / MLData EngineeringLLM SystemsSoftware Engineering

View Work GitHub LinkedIn

Selected Work

Data-driven AI systems and production software — from on-device inference to ETL pipelines serving real users.

On-Device AI · React Native · LLM Systems

Zayla

Active

React NativeTypeScriptllama.cpp (llama.rn)GGUFONNX Runtime

An offline-first mobile AI assistant that runs LLM inference locally on constrained Android hardware — zero servers, zero accounts, zero data collection.

System flow

01Mobile UI

02Model Manager

03GGUF Model

04llama.cpp

05CPU / GPU

06Local Inference

07Response UI

App screenshot

Metrics

Model: User-selected GGUF (device-scored)
Quantization: Q4_K_M / Q5_K_M
RAM usage: Model-dependent — scored pre-download
Performance: Benchmarking in progress (live in-app llama-bench)
Data collection: Zero — fully offline

Engineering challenge

Running useful LLM inference locally while staying inside mobile memory, thermal and compute constraints — without a server to fall back on.

Case Study GitHub

On-Device AI · LLM Systems · Low-Resource NLP

Africa LLM (AFRI-LLM-CODEX)

Active

Pythonllama.cppGGUFGBNF GrammarsNLLB-200

A fully offline coding assistant built for the hardware Africa actually has — CPU-only, 8GB RAM, zero network calls — submitted to the Africa Deep Tech Challenge 2026.

System flow

01Codebase

02DS-Code Graph (Context)

03Quantized LLM (GGUF)

04Grammar-Constrained Decoding

05JSON Patch / Explanation

Debug / Patch Mode

Interactive Mode

Pedagogical Tutor Mode

App screenshot

afri-llm-codex — run_inference.py

$ python run_inference.py --prompt "Fix the memory leak"

> peak RAM: 1.26 GB · 10.71 tok/s (CPU)

> arc_easy accuracy: 0.68 (n=50)

> grammar-constrained JSON patch, 100% offline

Q5_K_M

Q4_K_M

IQ4_XS

Metrics

Model: Qwen2.5-Coder-1.5B (Q5_K_M, 1.78B params)
Throughput: 10.71 tok/s (CPU)
Peak RAM: ~1.26 GB
Accuracy: 0.68 arc_easy (n=50)
Languages: 8, chrF++ 22.6–58.0

Research question

Can a single quantized model plus a from-scratch dependency-graph context engine deliver reliable, grammar-constrained code assistance entirely offline, on hardware with no reliable power or internet?

Case Study Competition

Data Engineering · ETL · Full-Stack Platform

Smart Billing System

Completed

PythonPostgreSQLFastAPIDjangoReact

A full-stack billing platform built around a high-volume ETL pipeline — ingesting, validating and serving 1.2M+ transaction records through a production-shaped API and dashboard.

System flow

01Transaction Data

02ETL Pipeline

03PostgreSQL

04REST API

05React / Streamlit UI

App screenshot

smart-billing — etl_pipeline.py

$ python etl_pipeline.py --source billing_raw --target postgres

> ingested 1,200,000+ transaction records

> processing speed: +34% vs. baseline

Baseline

Optimized

Metrics

Records processed: 1.2M+ transactions
Processing speed: +34% vs. baseline
Backend: FastAPI + Django
Database: PostgreSQL
Frontend: React + Streamlit

Engineering angle

How do you take a pipeline built to move 1.2M+ transaction records and turn it into a system real users — billing, customer management, payments — can depend on?

Case Study

Full-Stack Web · EdTech · Ongoing Maintenance

Mahder Academy

Active

Next.jsNode.jsMongoDBAuth & PaymentsCI/CD

A live e-learning platform I co-founded and built end-to-end — now serving 100+ students across 11+ courses, with ongoing maintenance and support.

System flow

01Student

02Next.js App

03API / Auth Layer

04Database

05Course Delivery

App screenshot

mahderacademy.com

Think Better. Speak Smarter.

Explore Courses →

Intermediate

Enroll →

Upper-Intermediate

Enroll →

Beginner (Kids)

Enroll →

Metrics

Role: Co-founder, full-stack build + maintenance
Students: 100+
Courses: 11+
Status: Live, in production
Support: Ongoing (solo)

Engineering angle

What does it take to keep a real production education platform reliable — not just launched — across new features, content changes and everyday support requests?

Case Study Visit Site

Earlier Work

Computer Vision · Deep Learning · Agri-Tech

Plant Disease & Fruit Ripening Detection

An earlier computer-vision project — detecting plant disease from leaf imagery and estimating fruit ripeness from visual cues — for greenhouse monitoring.

CNNOpenCVTransfer LearningImage Classification

Notes GitHub →

Engineering Depth

Grouped by discipline — only technologies I've actually shipped with.

AI / Machine Learning

Python
PyTorch
TensorFlow
scikit-learn
XGBoost
OpenCV
NumPy
Pandas

Data Engineering

Python
SQL
PostgreSQL
ETL / ELT Pipelines
Azure
AWS
GCP
Docker

LLM Systems

LangChain / LangGraph
RAG
LLM Fine-Tuning (QLoRA)
Prompt Engineering
llama.cpp
GGUF
Quantization
On-Device Inference

Software Engineering

TypeScript
React
React Native
FastAPI
Django
Node.js
Git
Linux

Experience

Co-founder & Full-Stack Developer
2024 — Present
Mahder Academy · Remote
- Designed and built a live e-learning platform end-to-end — frontend, backend API, and database — now serving 100+ students across 11+ courses.
- Own the platform after launch: uptime, bug fixes, and new features, working solo against real user feedback and usage data.
Next.jsNode.jsMongoDBFull-Stack
Software & Data Infrastructure Engineer
Jun 2024 — Apr 2025
Eritel Telecom · Asmara, Eritrea
- Engineered automated ETL pipelines in Python and SQL to process large-scale industrial datasets, increasing downstream data reliability by 28% and eliminating manual transformation steps.
- Architected scalable data models on Azure with high-concurrency support, enabling AI-assisted internal reporting tools used across multiple business units.
- Designed and shipped automated data validation and integrity checks that cut manual troubleshooting time by 33%, owning the feature from spec through production monitoring.
- Refactored high-volume datasets into query-optimised schemas, reducing query response times for data-heavy applications.
PythonSQLAzureETLData Modelling
Full-Stack & Analytics Engineer
May 2023 — Feb 2024
Eritrea Electric Corporation · Asmara, Eritrea
- Built and shipped full-stack dashboards and reporting pipelines end-to-end, enabling real-time visualisation of power distribution data for grid operations teams.
- Applied statistical modelling and trend analysis (deterministic and stochastic methods) to improve demand-forecasting accuracy by 23%, supporting grid capacity planning.
- Designed clean data architectures and automated processing scripts as the foundation for later machine-learning initiatives.
- Analysed large-scale consumption patterns, contributing to a 14% reduction in operational inefficiencies.
PythonForecastingData PipelinesDashboards
Research Assistant, Machine Learning
2022 — 2023
RIC Lab, Dept. of Computer Science and Engineering · Asmara, Eritrea
- Developed named-entity recognition (NER) and part-of-speech (POS) tagging resources for Tigrinya and Ge'ez — two under-resourced Semitic/Ethiopic languages with minimal prior NLP tooling.
- Designed annotation schemes and tagging pipelines addressing the near-total absence of labelled data, laying groundwork later extended into independent NLP projects.
NLPLow-Resource LanguagesNERAnnotation

Research

Low-resource NLP for Tigrinya and Ge'ez, and applied LLM experimentation — with links to the actual repositories.

Low-Resource NLP for Tigrinya & Ge'ez

Research Project

2022 – 2023 · RIC Lab, Dept. of Computer Science and Engineering

Named-entity recognition and part-of-speech tagging for two under-resourced Ethiopic-script languages with almost no prior NLP tooling.

Designed annotation schemes and tagging pipelines from a near-total absence of labelled data
Later extended with weak supervision and transformer-based sequence labeling

Repository

Tigrinya Tokenization (BPE)

Repository

Ongoing

A byte-pair-encoding tokenizer built specifically for Tigrinya, reducing token fragmentation and out-of-vocabulary rates for downstream NLP pipelines.

Repository

Multilingual Health QA — Zindi Competition

Experiment

2025

Fine-tuned CohereForAI/aya-23-8B via two-phase QLoRA (Kaggle, T4×2 GPUs) for multilingual health question answering under a fixed compute budget.

First submission scored 0.6559 ROUGE-L

Agentic RAG Research Assistant

Repository

2025 – Present

A LangGraph pipeline using multi-step tool calls and vector-store retrieval, with an evaluation loop to check answer quality systematically rather than trusting model output blindly.

Repository

About

I'm an AI and Data Engineer with a software engineering background, currently pursuing an MSc in Artificial Intelligence at the University of Jyväskylä.

I build data pipelines, machine learning systems, and AI products, with recent work focused on low-resource NLP, on-device LLM inference, and privacy-first AI applications.

My engineering work spans Python, SQL, Azure, PyTorch, React Native, Docker, and llama.cpp.

Trajectory

Software Engineering
Machine Learning Research
Data Engineering
LLM Systems
AI & Data Engineering

Contact

Get in Touch

GitHub LinkedIn Email

Haben Eyasu

Selected Work

Zayla

Africa LLM (AFRI-LLM-CODEX)

Smart Billing System

Mahder Academy

Plant Disease & Fruit Ripening Detection

Engineering Depth

Experience

Co-founder & Full-Stack Developer

Software & Data Infrastructure Engineer

Full-Stack & Analytics Engineer

Research Assistant, Machine Learning

Research

Low-Resource NLP for Tigrinya & Ge'ez

Tigrinya Tokenization (BPE)

Multilingual Health QA — Zindi Competition

Agentic RAG Research Assistant

About

Get in Touch