Kwabena Obeng

AI/ML && Software Engineering

Accra, Ghana — 5.6037° N, 0.1870° W

03 / Research

Research, Publications & Models

Applied research at the point where machine learning meets a constraint that actually bites — water quality sensed by ESP32 boards on an unreliable network, translation for a language with under ten thousand parallel sentence pairs, disease prediction from patchy public-health records. Published models are open on Hugging Face; the thesis and competition write-ups are below.

Research Experience

African Institute for Development Research and Evaluation

Jan 2024 - Sep 2024

Research Assistant · Remote

Study

Will Cameras Change Behaviours in Africa? The Case of CCTV Cameras on Streets in Ghana

  • Leveraged Python for extensive data analysis, utilizing criminal and conviction data from the Ghana Police Service
  • Conducted survey design and employed advanced data analysis techniques to extract valuable insights from responses
PythonData AnalysisStatistical ModelingSurvey DesignPolicy Research

Publications

Undergraduate Thesis — 2024

AQUAMIND: A Predictive Model for Water Quality Using Machine Learning and IoT Sensors

Developed an integrated IoT-ML system for real-time water quality prediction and monitoring. The system combines Long Short-Term Memory Neural Networks (LSTM), Gradient Boosting, and Random Forest algorithms to predict water quality based on critical parameters including pH, temperature, dissolved oxygen, turbidity, and biochemical oxygen demand (BOD). Integrates ESP32 microcontrollers with IoT sensors for real-time data collection, transmitted via HTTP requests to a Next.js web application. Achieved exceptional accuracy of 96.99% with LSTM and Random Forest, and 96.72% with Gradient Boosting.

Supervisor — Bright Yeboah-Akowuah, PhD

Department of Computer Engineering, Kwame Nkrumah University of Science and Technology

View thesis
LSTMRandom ForestGradient BoostingIoTESP32Next.jsWater Quality Prediction

Models

twi-en-nllb-v2 — Twi→English translation

Hugging Face
huggingface.co/ninte/twi-en-nllb-v2

NLLB-200-distilled-600M (615M parameters) fine-tuned for Twi (Akan) → English on the GhanaNLP parallel corpus (9,666 pairs). Reaches 27.30 BLEU and 51.03 chrF on a 967-pair held-out test set — a +16.8 BLEU improvement over the MarianMT v1 baseline. CC BY-NC 4.0.

Usage

python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "ninte/twi-en-nllb-v2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

tokenizer.src_lang = "twi_Latn"
inputs = tokenizer("Ɛsɛ sɛ mmofra di aduan pa na ama wɔanyin yiye.", return_tensors="pt")
out = model.generate(**inputs,
    forced_bos_token_id=tokenizer.convert_tokens_to_ids("eng_Latn"),
    num_beams=4, max_length=128, no_repeat_ngram_size=3)
print(tokenizer.decode(out[0], skip_special_tokens=True))
# → "Children need to eat healthy foods to grow well."
NLLB-200transformersLow-Resource MTTwifp16

twi-en-marianmt — Twi↔English baseline

Hugging Face
huggingface.co/ninte/twi-en-marianmt

MarianMT (Helsinki-NLP/opus-mt-mul-en) fine-tuned bidirectionally on 10,370 GhanaNLP Twi-English pairs as a v1 baseline: 10.50 BLEU / 32.30 chrF Twi→English on held-out data. The model card documents the failure modes — mixed-direction training without language tags, repetition loops, dataset noise — that shaped the NLLB v2. CC BY-NC 4.0.

Usage

python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_id = "ninte/twi-en-marianmt"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

inputs = tokenizer("Ɛsɛ sɛ mmofra di aduan pa.", return_tensors="pt")
out = model.generate(**inputs, num_beams=4, max_length=128,
                     no_repeat_ngram_size=3, early_stopping=True,
                     length_penalty=1.0)
print(tokenizer.decode(out[0], skip_special_tokens=True))
MarianMTtransformersLow-Resource MTTwi

Research Interests

Machine Learning, AI in Healthcare, Computer Vision & Robotics, Deep Learning, LLMs & NLP, Scalable AI, AI Ethics, and Responsible AI Development for Social Good.

Achievements

SUA Outsmarting Outbreaks Challenge

SilverZindi Africa

Can you turn the tide on waterborne diseases by predicting the next outbreak in Tanzania?

Developed machine learning models to predict outbreaks of climate-sensitive waterborne diseases (typhoid, amoebiasis, diarrhoea, schistosomiasis, intestinal worms) in Tanzania using comprehensive datasets spanning water sources, sanitation quality, waste management, health facilities, and climate data from 2019-2023.

Impact

Enabling governments and health organizations to implement timely, targeted interventions and optimize resource allocation for vulnerable populations.

PythonMachine LearningTime Series AnalysisPublic Health AnalyticsClimate Data Analysis

Ghana Crop Disease Detection Challenge

BronzeZindi Africa

Can you build a model for mobile phones to identify diseases on tomatoes, corn, and peppers?

Built robust machine learning models for accurate prediction of multiple diseases in corn, pepper, and tomato crops, with focus on generalization to unseen diseases and efficient operation on entry-level smartphones used by subsistence farmers in Africa.

Impact

Supporting food security for millions by enabling timely disease detection in crops that form the backbone of Sub-Saharan African agriculture.

Computer VisionMobile MLEdge ComputingAgricultural AIDisease Classification

Kenya Clinical Reasoning Challenge

SilverZindi Africa

Can your model match real clinicians in rural Kenyan healthcare?

Developed AI models to replicate clinical reasoning of trained healthcare professionals using 400 authentic clinical prompts from diverse Kenyan healthcare settings. Models predict clinician responses across maternal health, child health, and critical care scenarios.

Impact

Supporting frontline healthcare workers in resource-limited settings with AI-assisted clinical decision making.

NLPClinical AIHealthcare AnalyticsSmall Data MLMedical Reasoning