Skip to content

AI Engineer & Researcher

Haben Fessehaye

LLMs · NLP · Computer Vision · Efficient AI

I build and investigate AI systems, with a particular interest in low-resource languages, efficient inference, and practical machine learning.

View Selected WorkResearch InterestsGitHub
inputhiddenoutput

Selected Work

Selected Work

A few projects that represent how I think, build and experiment with AI.

LLM · Edge AI · Efficient Inference

Offline African-Language LLM

Active
CPU-only8GB RAMGGUFllama.cppGrammar-constrained generation

Built an offline LLM system designed to operate on commodity hardware while supporting African-language interaction under strict compute and memory constraints.

System flow

01User
02Agent Orchestrator
03Quantized LLM
04Grammar Constraint
05Tool / Response

Metrics

Model
Qwen2.5-1.5B-Instruct
RAM
≤ 8 GB
Tokens/sec
8–12 tok/s (CPU)
Model Size
~1.1 GB (Q4_K_M)
Accuracy
In evaluation

Research angle

How do model size, quantization, grammar constraints and hardware limitations affect accuracy and inference efficiency?

Case StudyGitHub

NLP · Low-Resource AI · Research

Tigrinya NLP

Active
TokenizationGe'ez ScriptNERLanguage ModelingPredictive Text

Exploring practical NLP methods for Tigrinya through tokenization, named entity recognition, predictive text and language modeling.

System flow

01Tigrinya Text
02Tokenization
03Representation
04NER / Language Modeling
05Prediction
Tokenizer
Ge'ez NER
Predictive Keyboard
Language Modeling

Metrics

Script
Ge'ez (ልቢ)
Speakers
~9M
Corpus size
Actively growing
Tokenizer fertility
Under evaluation
NER F1
Baseline in progress

Research angle

How can NLP systems become more useful for languages with limited linguistic and training resources?

Explore ProjectGitHub

Computer Vision · Pose Estimation · Mobile AI

Mobile Jump Height Estimation

Active
Pose EstimationMediaPipeMonocular CVBiomechanics

Investigating whether smartphone cameras can estimate vertical jump height using pose estimation and biomechanical signals.

System flow

01Camera
02Pose Detection
03Keypoints
04Motion Analysis
05Jump Height Estimation

Metrics

Input
Single monocular camera
Keypoint model
MediaPipe Pose
Method
Flight-time kinematics
Frame rate
30–60 fps
Accuracy
Benchmarking in progress

Research question

Can consumer smartphone cameras estimate vertical jump height reliably enough for practical athletic applications?

Case StudyGitHub

Computer Vision · Deep Learning

Plant Disease Detection

An earlier deep-learning project exploring computer vision for greenhouse monitoring and plant disease detection.

CNNImage ClassificationGreenhouse Monitoring
NotesGitHub →

Research

Research

Questions I am interested in exploring at the intersection of machine learning, language and real-world constraints.

01

Low-Resource NLP

How can NLP systems become more useful for languages with limited datasets and linguistic resources?

02

Efficient LLMs

How can useful language models operate under strict memory, compute and latency constraints?

03

Computer Vision

How can vision systems reliably estimate physical properties from consumer devices?

04

Responsible AI

How can we evaluate whether AI systems behave reliably across different populations and contexts?

Currently exploring

  • Low-resource NLP
  • Efficient LLM inference
  • Multimodal AI
  • AI evaluation

Research & Experiments

Research & Experiments

Not everything begins as a finished product. Some ideas begin as questions.

Grammar-Constrained LLM Agents

Active

Exploring structured generation and tool calling using formal grammars.

GitHub →

Human Information Loss & Prompting

Exploring

Exploring whether the way humans compress experiences before communicating them affects downstream LLM performance.

African-Language NLP

Active

Investigating practical approaches to NLP in Tigrinya and related low-resource settings.

GitHub →

AI Evaluation & Bias

Exploring

Exploring how model assumptions and evaluation choices affect reliability across populations and contexts.

Engineering

From Research to Working Systems

I don't stop at the model. I build the system around it.

Research
Model
Inference
Backend
Application
Deployment

Tools I build with

Languages

  • Python
  • Java
  • JavaScript / TypeScript

AI / ML

  • PyTorch
  • TensorFlow
  • Hugging Face

NLP / LLM

  • Transformers
  • llama.cpp
  • GGUF
  • Quantization
  • RAG
  • LLM evaluation

Computer Vision

  • CNNs
  • Pose Estimation
  • MediaPipe

Engineering

  • React
  • Node.js
  • Firebase
  • MongoDB
  • Linux

Experience

Experience

  1. ML Engineer

    2025 — Present

    Eritrean Electric Corporation

    • Design and ship machine-learning components for infrastructure and operations use cases.
    • Focus on model reliability, evaluation, and turning prototypes into systems that run in production.
    Machine LearningProduction SystemsEvaluation

  2. ML Researcher

    2024 — Present

    Independent / Applied Research

    • Lead experimentation on low-resource NLP, efficient LLM inference, and computer vision problems.
    • Frame open questions as testable experiments and document constraints, methods and results.
    NLPLLMsComputer VisionResearch

  3. AI Data / Model Evaluation

    2023 — 2024

    Contract & Applied Work

    • Worked on data annotation, model evaluation, and quality/instruction review for AI training pipelines.
    • Supported model training with structured feedback loops between data quality and model behavior.
    Data AnnotationAI EvaluationModel Training Support

  4. Graduate Assistant

    2023 — 2024

    University

    • Taught and supported coursework in Object-Oriented Programming, Java, Data Structures, and Python.
    • Held office hours and labs translating core CS fundamentals into practical programming skill.
    TeachingJavaData StructuresPython

  5. Software Engineer

    2021 — 2023

    Early Career

    • Built full-stack applications and backend systems — the engineering foundation behind current AI work.
    • Developed the systems instincts now applied to shipping models as real, usable products.
    Full-StackBackendSystems

Thinking About AI

Thinking About AI

Notes, experiments and ideas from the lab.

All writing →
Efficient AI6 min

Can an 8GB laptop run a useful LLM?

A practical look at what 'useful' has to mean once you fix the hardware budget at 8GB of shared RAM and no GPU.

Jul 12, 2026

Structured Generation5 min

Why grammar constraints matter for small language models

Small models are unreliable narrators. Grammar-constrained decoding turns that unreliability into a solved problem for structured output.

Jun 3, 2026

Low-Resource NLP7 min

What makes Tigrinya NLP difficult?

Morphology, script, and missing benchmarks: an honest inventory of why Tigrinya doesn't fit off-the-shelf NLP tooling.

Apr 21, 2026

Computer Vision5 min

Can a smartphone camera measure jump height?

A single monocular camera, no depth sensor, and a question borrowed from biomechanics: how far can flight-time kinematics get you?

Mar 8, 2026

Prompting6 min

What happens when humans compress information before prompting an LLM?

People don't type what they mean — they type a compressed summary of it. What does that lossy compression do to downstream model performance?

Jan 19, 2026

About

About

I'm an AI engineer and researcher interested in building intelligent systems for environments where conventional AI assumptions break down — limited compute, limited data, and underrepresented languages.

My work spans NLP, LLMs, computer vision and software engineering, with a particular interest in African languages and efficient AI.

Trajectory

  1. Software Engineering
  2. Machine Learning
  3. Computer Vision
  4. NLP
  5. Low-Resource AI
  6. LLMs + Efficient AI
  7. Research

Contact

Let's build or investigate something interesting.

Interested in AI research, applied machine learning, low-resource NLP or efficient AI?

EmailLinkedInGitHub