Available now · full-time DS / AI roles

Fernado George · Data Scientist & AI Engineer

Models that leave the notebook.

I'm a data scientist and AI engineer in Malaysian Borneo. I build retrieval assistants, computer-vision pipelines and the interfaces that put them in people's hands, then measure whether they hold up outside the lab.

See selected work Resume

multimodal_fusion.infer()

true label sad

P(predicted)

Row-normalised confusion matrix from the thesis. Pick a label to compare.

Thesis F1 (multimodal): 71.6%
CGPA · Dean's List: 3.78
Client systems in production: 2
Featured hackathon builds · 1 Gold, 1 runner-up: 4

Python
Java
C++
R
SQL
SQLAlchemy
JavaScript
HTML/CSS
Deep Learning
NLP
LLMs (OpenAI, Gemini, Ollama)
RAG (FAISS, BM25)
Computer Vision
LightGBM & Forecasting
PyTorch
TensorFlow
FastAPI
React (Vite)
Node.js
Tailwind CSS
Gradio & Streamlit
LangGraph & CrewAI
MLflow
Docker & Kubernetes
PostgreSQL, MySQL & MongoDB
AWS (ECS, ECR, S3)
Git & CI/CD
Python
Java
C++
R
SQL
SQLAlchemy
JavaScript
HTML/CSS
Deep Learning
NLP
LLMs (OpenAI, Gemini, Ollama)
RAG (FAISS, BM25)
Computer Vision
LightGBM & Forecasting
PyTorch
TensorFlow
FastAPI
React (Vite)
Node.js
Tailwind CSS
Gradio & Streamlit
LangGraph & CrewAI
MLflow
Docker & Kubernetes
PostgreSQL, MySQL & MongoDB
AWS (ECS, ECR, S3)
Git & CI/CD

01/Selected work

Things I've built and shipped

9 builds across retrieval, document AI, computer vision, forecasting and social-impact products, including two client systems running in production.

AI & HealthcareFeatured
Gold Medal · Festival Idea Sabah 2025
Team Leader of Development
Built an AI medical scribe with Google Speech-to-Text and Gemini Pro 2.5, using all-MiniLM-L6-v2 embeddings in a RAG pipeline that grounds generated notes in retrieved context.
- Google Speech-to-Text
- Gemini Pro 2.5
- RAG
- GCP
Client project: screenshots and repository are private.
Retrieval & LLM Systems
Client work, in production
AI/ML Development Intern · Breakfast Byte
Gradio-based AI assistant for a German client, running local Ollama models alongside OpenAI workflows. Hybrid BM25 + FAISS retrieval over a multi-format document pipeline (PDF, DOCX, spreadsheets) with streaming responses, so staff query internal documents in natural language instead of searching folders.
- Python
- Gradio
- +4
Client project: footage and repository are private.
Computer Vision
Client work, in production
AI/ML Development Intern · Breakfast Byte
End-to-end vision pipeline converting a Sarawak car-wash operator's daily CCTV footage into structured service records: vehicle-session detection, multi-bay assignment, license-plate OCR, and wash-action classification, surfaced through a local operator dashboard.
- Python
- OpenCV
- +3
NLP & Social Impact
2nd Prize · JomHack Varsity Challenge 2025
Lead Developer
AI-powered matchmaking engine connecting under-resourced schools with NGOs. A Telegram-first intake flow turns a plain-language description of need into a structured request, then ranks verified donors against it. There's no web portal for schools to learn.
- AI Chatbot
- +3
Full-Stack Web App
Event complete · 60+ participants · 100% uptime
Full-Stack Developer
Handled real-time registration for 60+ international participants with 100% uptime across the event weekend.
- React (Vite)
- Node.js
- +2
AI & Marine Tech
Top 8 Finalist · MarineHack 2025
AI/ML Engineer
Predictive maintenance system that flags a hull's biofouling growth before drag builds, so cleaning gets scheduled ahead of the fuel-cost spike instead of after it. The growth model is paired with A* route optimization. On simulated voyage data the combined model projected up to 30% fuel savings (a simulation result, not a sea trial).
- FastAPI (Python)
- +3
Social Impact & AI
Borneo Hackathon 2025
Team Leader & Backend Developer
'Zero-tech' AI coaching app for rural micro-entrepreneurs: a localized 'Cikgu' chatbot speaking Bahasa Pasar, photo-to-copywriting generation, and daily missions like 'log today's sales in one voice note'. It's built for owners whose only device is a shared phone.
- React Native
- +3
Data Engineering & MLOps
End-to-end telemetry platform · Docker & Kubernetes
Solo Developer
Renewable-energy telemetry platform ingesting live Open-Meteo data into a PostgreSQL star-schema warehouse with a DuckDB OLAP layer and n8n orchestration. LightGBM inflow forecasting and Isolation Forest anomaly detection (MLflow-tracked) are served via FastAPI, with a LangGraph self-correcting SQL agent, a CrewAI strategic panel, and a Streamlit SCADA command center. Everything is containerized with Docker Compose and deployable to Kubernetes.
- Python
- PostgreSQL
- +9
Document AI & Accounting
311 backend tests · audited and re-verified
Solo Developer
Receipt workspace for the kind of business that runs on thermal paper: phone photos in, a FastAPI and React review console out. EasyOCR keeps normalized bounding boxes, so every extracted field traces back to the pixels it came from, and an LLM stage structures the raw tokens. The reconciler is tax-basis aware, which is the interesting part: Malaysian receipts print GST-inclusive line amounts, and the first version assumed exclusive, so every GST receipt failed the arithmetic check and the export carried the wrong tax basis. Found by driving the app end to end, fixed, and written up in a public audit report. Exports a purchase-journal CSV and a MyInvois-shaped JSON draft, the latter never submitted to LHDN.
- Python
- FastAPI
- +6

Open any project for the full write-up · ← → to browse

02/Final Year Thesis

Does emotion recognition survive the real world?

Addressed the gap between lab performance and real-world application in Emotion Recognition. This study benchmarked Contextual vs. Statistical fusion strategies against noise and missing data, concluding that Ensemble methods offer the highest reliability while Contextual Bi-GRU offers the best efficiency.

Comparative Study of Hybrid Fusion for Robust Multimodal Emotion Recognition

Weighted F1, ensemble: 71.6%
Models benchmarked: 15
Modalities fused: 3
Emotion classes: 6

01
Robustness benchmark. Every model re-scored under Gaussian noise and dropped modalities, not just clean input.
02
Contextual Bi-GRU. Architecture capturing temporal emotion changes better than statistical baselines.
03
Ensemble strategy. Achieved 71.6% F1-score, maintaining stability even under noise.

Python
PyTorch
WavLM
DINOv2
ModernBERT

Live presentation Read the thesis

Fig. 01

What happens when the inputs get noisy

y: weighted F1 %

Weighted F1 as Gaussian noise is added to every input. Each band spans the best and worst model in a family. Values above σ = 0 are read off the thesis plots, so treat them as approximate.

03/Experience

Where the work happened

A completed six-month AI/ML internship with two client systems in production, plus two terms on my faculty's student association committee.

Mar 2026 – Aug 2026

Completed

AI/ML Development Intern

Breakfast Byte Sdn. Bhd. · Kuching, Sarawak

Six-month AI/ML internship, now completed. Two client systems in production: KI Konnekt, a hybrid BM25/FAISS retrieval assistant for a German client, and a CCTV-to-records computer-vision pipeline for a Sarawak car-wash operator, both written up in Projects. Also built HydroGrid Sarawak, an open-source renewable-energy telemetry platform: PostgreSQL star-schema warehouse with DuckDB OLAP, MLflow-tracked forecasting and anomaly detection served via FastAPI, and LangGraph/CrewAI agents deployed via Docker and Kubernetes.

Built during the internship

KI Konnekt / client RAG assistantRetrieval & LLM Systems
CCTV-to-records pipeline / car washComputer Vision
HydroGrid Sarawak / Energy & Hydrogen IntelligenceData Engineering & MLOps

Student leadership

Nov 2023 – Aug 2024
Ketua Exco Akademik dan KerjayaHead of Academic & Career Committee
Persatuan Mahasiswa Fakulti Komputeran dan Informatik (PMFKI) · Universiti Malaysia Sabah
Held the academic and career portfolio for the computing faculty's student association, running its academic programmes and industry engagement for a 10-month term.
Nov 2022 – Aug 2023
Timbalan SetiausahaDeputy Secretary
Universiti Malaysia Sabah
Deputy secretary for the association: meeting records, committee correspondence and event documentation across a 10-month term.

Education

Bachelor of Computer Science (Data Science) With Honours
Universiti Malaysia Sabah · 2022 – 2026 · Graduated
CGPA 3.78 · Dean's List
Science Module II
Sarawak Matriculation College · 2021 – 2022
CGPA 4.00 · Best Computer Science Student Award

Toolkit

Languages8

Python
Java
C++
R
SQL
SQLAlchemy
JavaScript
HTML/CSS

AI & ML6

Deep Learning
NLP
LLMs (OpenAI, Gemini, Ollama)
RAG (FAISS, BM25)
Computer Vision
LightGBM & Forecasting

Frameworks & Tools13

PyTorch
TensorFlow
FastAPI
React (Vite)
Node.js
Tailwind CSS
Gradio & Streamlit
LangGraph & CrewAI
MLflow
Docker & Kubernetes
PostgreSQL, MySQL & MongoDB
AWS (ECS, ECR, S3)
Git & CI/CD

04/About

The short version

Fernado George Anak Mani

I started out writing small game mods, which turned into a habit of taking systems apart to see how they decide things. That is still what I do, now with retrieval systems and computer vision.

My thesis benchmarked multimodal emotion models under noise and missing inputs, reaching 71.6% F1. My internship work turned a client's document archive into a queryable assistant and their CCTV feed into structured service records. I build the model and the interface that puts it in someone's hands.

Off-hours, I fish.

Based in: Malaysian Borneo
Focus: Retrieval systems · Computer vision
Education: Bachelor of Computer Science (Data Science) With Honours, Universiti Malaysia Sabah
Grades: CGPA 3.78 · Dean's List

Hackathons

Four hackathons (MarineHack, JomHack, Festival Idea Sabah and Borneo Hackathon) with one Gold, one second place and one top-8 finish (MarineHack), plus a second top-8 at the ELLM Startup Initiative. Roles ranged from leading the team and owning the backend to solo full-stack delivery.