Natnael Fisseha
I build end-to-end AI systems, from machine learning pipelines and LLM-powered applications to the backend infrastructure and interfaces that make them work in production.
FromAIResearchtoProductionSystems
I'm a Software Engineering graduate from Addis Ababa Science and Technology University (AASTU), currently completing an MSc in Software Engineering. My work spans machine learning engineering, applied AI, and full-stack backend development.
At Gheero, I built production ML systems end to end: an early-warning classifier for task overdue prediction, a hybrid recommendation engine matching trainees to jobs using embeddings and skill scoring, a dynamic task allocation system for a live crowdsourcing platform, and a low-latency Amharic speech-to-speech pipeline with no external API dependencies. I also conduct research on RL-based tokenization for Amharic and data-efficient generative model training.
Alongside this, I run an independent consulting practice building AI-powered systems for real clients, with a 5.0 rating across completed Upwork contracts.
I am particularly interested in:
WhereI'veWorked
ML Engineer
Gheero AI Residency
- Optimized an Amharic banking voice assistant (PyTorch, llama.cpp, OmniVoice), reducing time-to-first-audio by 85% (9.02s to 1.2s) through asynchronous streaming, 4-bit quantization, and multi-GPU inference; increased LLM throughput 2.8x from 18.6 to 53 tok/s
- Built RAG and hybrid recommendation systems (FastAPI, pgvector, MiniLM, Gemini API) over a 650+ skill taxonomy; achieved 48.1ms average query latency, 0.88 Precision@5, and 1.00 Hit Rate@5
- Segmented 884,265 retail banking customers using K-Means, GMM, autoencoders, and PCA; validated a 3-cluster solution with bootstrap ARI 0.80 and 4-cluster hierarchical drill-down with 99.5% nesting stability
- Trained an RL-based Amharic morphological tokenizer (REINFORCE, 29,867-verse corpus); achieved 0.952 morphological alignment at a 3,000-token vocabulary, outperforming BPE, WordPiece, and SentencePiece by up to 27 points
- Built task-suitability and overdue-risk classifiers (CatBoost, XGBoost) achieving 0.857 and 0.712 ROC-AUC under leakage-safe forward-chaining chronological evaluation
- Published empirical research on knowledge distillation for small language models (7.24M to 19.22M parameters), examining teacher-student domain transfer across the TinyStories corpus.
Applied AI Engineer
Independent Consultant
- Designed and deployed LLM-powered applications using RAG, embeddings, prompt engineering, and agentic workflows for clients across education, real estate, and business services
- Built AI-driven educational systems with automated question generation, scoring pipelines, and curriculum-aligned RAG feedback
- Built a voice-based travel planner integrating ElevenLabs, n8n, and LLM reasoning to generate and deliver personalized itineraries
- Maintained a 5.0 rating across all Upwork contracts; clients highlighted production-ready delivery, error handling, and treating projects like their own product
Available for contract hire & AI systems engineering
Backend Engineer
Dynamic Technologies
- Migrated a live production platform to an in-house Node.js and Hono.js architecture, reducing response latency by 80% with zero downtime
- Designed PostgreSQL schemas, indexing strategies, and real-time data ingestion pipelines under a two-week delivery constraint
- Built and containerized microservices with Docker and Kubernetes; enforced JWT auth, RBAC, rate limiting, and input validation across all endpoints
Technical Lead and Mentor
GDG on Campus AASTU
- Led a Node.js and AI/ML development track with 7 mentors and 70+ developers
- Designed curriculum, coordinated capstone sprints across 15 teams, and ran structured technical evaluations
- Facilitated sessions on AI/ML engineering, software architecture, and collaborative development
AISystemsI'veBuilt
Kelem.app
Multi-Tenant School Management SaaS
A full multi-tenant SaaS platform with isolated data per organisation and role-based dashboards for administrators, teachers, students, and parents. Covers attendance, assessments, assignments, scheduling, announcements, parent-teacher communication, and analytics.
- Multi-tenant PostgreSQL schema with full data isolation per school
- JWT access and refresh token auth with Redis caching for high-frequency reads
- Bulk data import pipeline for school onboarding
- Analytics dashboards for administrators and role-specific views for all user types
AI B2B Lead Generation and Outreach Platform (Neuro)
Automated Prospecting and Outreach
End-to-end automation replacing manual prospecting, handling up to 100 leads per campaign.
- Scrapes and enriches leads, scores prospects with Gemini, delivers personalized outreach via email and voice
- Round-robin email rotation, Telegram bot dashboard for campaign management
- AI-generated outreach websites built from each business's own data
- Outbound voice calls through Twilio and ElevenLabs
Automated Web Monitoring and Data Processing System
Production-Ready Monitoring System
Production-ready monitoring system built with Playwright and Node.js, including a scalable backend and a clean monitoring dashboard.
- Session handling, error recovery, structured logging, and secure role-based access
- Client review: "He delivered a production-ready solution... treated the project like his own product."
DMerang
Instagram Cold-DM Automation
Cold DM automation tool built to work within Instagram's policy limits, with lead scraping, scheduling by timezone, and multi-account support.
Amazon E-Commerce Test Automation
End-to-End UI Test Automation
UI test automation framework covering key purchase flows with Page Object Model architecture, anti-detection browser config, screenshot capture on failure, and CI/CD-ready HTML reporting.
Voice Travel Planner
Agentic Travel Assistant
A conversational travel planner that transforms natural voice dialogue into complete personalized itineraries delivered by email.
TechnicalResearch&SystemsStudies
Empirical ML investigations, model compression, low-resource morphological NLP, and autonomous self-healing agent architectures.
Off-the-shelf teachers do not improve validation loss at any student size. Domain mismatch between general-purpose pretraining and targeted student corpora is the primary driver.
Tools&TechnologiesIUse
ML and AI
Agentic Systems and Automation
Backend and Full-Stack
Infrastructure
Honors, Leadership & Certifications
Verified academic excellence, community leadership, and specialized technical accreditations.
Honors & Awards
HackAtom Ethiopia 2024 Winner
National WinnerNational hackathon, Rosatom & RUDN University partnership
View VerificationGraduated with Very Great Distinction
Academic TopHighest academic honor at AASTU
View VerificationAASTU Presidential Award
Conferred for outstanding institutional contributions
View VerificationDean's Award for Academic Excellence
School of Electrical & Mechanical Engineering
View VerificationHult Prize 2026 Semifinalist
Global social entrepreneurship competition
View VerificationGAIA 2025 Competition Participant
Global AI Agent benchmarking challenge
View Verification
Leadership & Community
Technical Lead and Mentor, GDG on Campus AASTU
70+ Devs MentoredMentored 70+ student developers across 7 specialized mentors and 15 capstone teams. Designed structured technical evaluation rubrics and hands-on workshops.
View VerificationCapstone Technical Evaluation Board
Conducted sprint reviews, code architecture assessments, and production readiness checks for collegiate software teams.
Professional Certifications
DataCamp Certifications (2024–2025)
5 Track CertificatesWorking with Hugging Face, Introduction to Embeddings with OpenAI API, LLMOps Concepts, Developing AI Systems with OpenAI, Prompt Engineering with OpenAI API
View VerificationFundamentals of Data Science
Cisco VerifiedCisco Networking Academy: Data science methodologies, data analysis pipelines, Python statistical modeling.
View Verification
Ready to Build Reliable AI Systems?
Open to AI/ML systems engineering, LLM application architecture, and high-performance backend roles. Available for direct client contracts and consulting via Upwork.
Accepting new consulting projects, AI system deployments, and full-time technical opportunities.
Location
Remote WorldwideAddis Ababa, Ethiopia (UTC+3)