Prince N. Alhassan

Building language technology for the languages of Northern Ghana.

Machine translation and speech recognition for Kusaal, a Mabia language of roughly 400,000 speakers in northern Ghana and southern Burkina Faso, and across 42 Ghanaian language subsets. Open models, an open corpus, a public benchmark, and results measured on held-out data.

↓ · University of Ghana, Legon · CGPA 4.0

Measured on held-out data.

Source-disjoint splits throughout: no speaker or passage appears in both training and test.

0.0

BLEU, Kusaal → English

0

Languages, one ASR model

0

Corpus pairs built

0.00

Hours of audio aligned

Prince Nasamu Alhassan is an 18-year-old AI researcher and software engineer at the University of Ghana, Legon, maintaining a 4.0 CGPA. A native Kusaal speaker from Bawku in Ghana's Upper East Region, he independently designed, trained, and publicly released an open-source Kusaal-English machine translation system — building the roughly 63,000-pair corpus that underlies it from scratch, most recently by mining 13,659 parallel sentences from Kusaal Wikipedia.

His work has been presented at a GhanaNLP community session, where GhanaNLP co-founder Paul Azunre publicly engaged and requested access to the dataset. He builds full-stack products used by real students, contributes to the Kusaal Wikimedia Community, and taught himself programming during COVID-19 lockdowns before any formal CS instruction.

Based in
Accra, Ghana
From
Bawku, Upper East
Studying
BSc Mathematical Sciences with Computer Science
CGPA
4.0 / 4.0
Graduating
2029
Languages
Kusaal, English

Languages & Frameworks

Python · JavaScript · TypeScript · React · Next.js 14

ML / NLP

PyTorch · HuggingFace Transformers · NLLB-200 · Whisper · w2v-BERT 2.0 · scikit-learn · RAG Systems · BLEU Evaluation

Infrastructure & Data

Supabase · Vercel · FastAPI · Flask · Git/GitHub · Kaggle · Jupyter · SQL · REST APIs · pandas

Research

Low-resource MT · Parallel Corpus Curation · Data Augmentation (Back-translation) · ASR Fine-tuning · Multilingual NLP

Human Languages

Kusaal (native) · English (professional)

Built from scratch, released for everyone.

The corpus behind the translation system was assembled pair by pair from six separate sources, then extended with back-translation and a sentence-aligned corpus mined from Kusaal Wikipedia — roughly 63,100 pairs, published under CC BY 4.0 and CC BY-SA 4.0.

The speech corpus is separate: 30,820 verse-level clips totalling 81.71 hours at 16kHz mono, verified at zero integrity errors.

Corpus provenance

~63,100 pairs

Bible text: 46.4%
Wikipedia harvest: 21.6%
Back-translation: 14.7%
English-Kusaal Index: 5.6%
GhanaNLP: 5.5%
Lexique Pro: 4.4%
Wikipedia (original): 1.8%
  • Bible text46.4%
  • Wikipedia harvest21.6%
  • Back-translation14.7%
  • English-Kusaal Index5.6%
  • GhanaNLP5.5%
  • Lexique Pro4.4%
  • Wikipedia (original)1.8%

Decisions that changed the numbers

  1. 01

    Seeding kus_Latn from Dagbani

    The new language token was initialised from the Dagbani embedding rather than randomly. Dagbani is the closest well-resourced Mabia relative, so the model starts from a related point in embedding space instead of noise.

  2. 02

    Splitting speech by book, not at random

    Random clip-level splits let the same narrator and the same passage appear in both train and test, which inflates the score. Splitting at book level keeps the evaluation honest.

  3. 03

    Testing narrator leakage across 42 languages

    Six languages in the corpus contain the same scripture read by different narrators, so a random split can put byte-identical text on both sides. The splits were rebuilt to be book-disjoint — and the cost of the leak was then measured directly and published as a negative result: on these weights, the difference sat within noise.

  4. 04

    Mining Wikipedia for a second register

    A shared-anchor alignment pipeline, verified by the model and partially human-adjudicated, mined 13,659 encyclopedic sentence pairs from Kusaal Wikipedia — lifting the weaker direction by 12.8 BLEU on encyclopedic text, with no forgetting on the original test set.

  5. 05

    Freezing a public benchmark

    A 1,000-pair test set from the Wikipedia corpus is frozen and published with its evaluation protocol and reference results, so anyone can verify the numbers independently — and no one, including the author, trains on it.

Projects

Research systems, shipped products and the engineering underneath them.

research2025 – present

Kusaal-English Machine Translation

An open-source machine translation system for Kusaal, a Mabia language spoken by roughly 400,000 people in northern Ghana and southern Burkina Faso. Fine-tuned NLLB-200-distilled-600M with the kus_Latn token seeded from Dagbani rather than randomly. The corpus — now ~63,100 pairs across seven sources — was extended in August 2026 with 13,659 sentence pairs mined from Kusaal Wikipedia, lifting encyclopedic-text BLEU to 47.73 (ks→en) and 32.26 (en→ks), a +12.8 BLEU jump in the harder direction, with no forgetting on the original test set. Ships with a frozen, public 1,000-pair benchmark so the numbers are independently verifiable.

39.2

BLEU ks→en

26.2

BLEU en→ks

63,100

Corpus pairs

NLP · HuggingFace · Low-resource MT · PyTorch · NLLB-200

research2026

Multilingual ASR for Ghanaian Languages

One speech model for 42 Ghanaian language subsets, trained in a 48-hour H200 window at the GhanaNLP TTS/ASR Hackathon. w2v-BERT 2.0 with a CTC head in bf16, building on GhanaNLP's DONDO baseline, plus a jointly trained language-ID head — speak, and it works out which of the 42 languages it heard, at 96.8% mean accuracy. Six languages contain the same scripture read by different narrators, so the splits were rebuilt to be book-disjoint; the cost of that leakage was then measured directly and published as a negative result rather than a scare number. 30.2% micro WER on a test set where no book appears in training.

30.2%

WER, book-disjoint

96.8%

Language ID accuracy

42

Language subsets

ASR · w2v-BERT 2.0 · CTC · Language ID · ONNX

product2026

Forge — AI-Powered Academic Productivity Tool

A voice-driven study scheduler, designed and shipped in 48 hours.

React · Vite · AI · Vercel · Hackathon

product2026 – present

CampusLink — Verified Student Connection Platform

Full-stack platform with two-stage student identity verification, real-time messaging, and an interactive campus map. Built for UG Legon students. Also in development as CampusLink Pay for Moolre Startup Cup 2026.

Next.js 14 · Supabase · Gemini API · Full-stack

product2026

Mabia AI — Maternal Health Voice Platform

An offline-first voice platform for CHPS maternal health work in Northern Ghana, built for the UNICEF StartUp Lab AI for Nurturing Care Hackathon 2026. It calls pregnant women and caregivers in their own language over plain GSM — triaging danger signs, measuring dietary diversity, and dispatching community transport. The transport map is drawn on time-from-care rather than coordinates, because no village driver in the system has coordinates; the screen that matters shows the community with women enrolled and no vehicle that can carry them.

Voice · FastAPI · Health · Offline-first · Hackathon

research2026

Kusaal ASR — Whisper Fine-tune

Eighty-one hours of audio, assembled by hand, into a Kusaal speech recogniser.

30.41%

Word error rate

81.71hours

Audio

30,820

Verse-level clips

ASR · Whisper · LoRA · Low-resource

engineering2026

Ghana AI-NCD Navigator

Five connected services for national NCD care, wired into a real DHIS2 instance.

Flask · DHIS2 · FHIR · Health Systems · Next.js 14

engineering2026

Zigi Voice — Local-Language Telco Assistant

Say what you need in Twi, Kusaal or English; it finds the product and stops at the PIN.

Voice UI · Twi · Kusaal · React · Dialogue Systems

product2026

Editorial Board Hub

A school editorial board, run end to end: public site, members’ portal, archive, SMS.

Next.js 14 · Postgres · Full-stack · SMS

product2026

FieldMind — RAG Document Intelligence

Retrieval-augmented search over enterprise document collections.

RAG · Gemini · LangChain · Hackathon

product2026

ShieldNet AI

Threat intelligence, mobile money and emergency SMS in one security platform.

AI · Cybersecurity · SaaS · In Development

engineering2026

HyperFlow Risk Agent

Multi-agent risk analysis for live trading decisions.

AI Agents · Risk Intelligence · AMD

engineering2026

WhatsApp & Telegram AI Agents

Locally-hosted agents that summarise WhatsApp and answer on Telegram.

Telethon · Ollama · Automation · Python

engineering2026

DecodeLabs Climate Data Pipeline

Climate time-series from five Ghanaian cities, ingested end to end.

Data Engineering · Python · Open-Meteo · Internship

engineering2026

VerifiQ

QR attendance tracking for lecture halls.

QR · Attendance · Web

Running in production

The real deployments, embedded as they are at this moment — not screenshots.

forge-study-theta.vercel.app
Open

Live preview available on a larger screen

Forge — AI-Powered Academic Productivity Tool

2026

A voice-driven study scheduler, designed and shipped in 48 hours.

campus-link-sigma.vercel.app
Open

Live preview available on a larger screen

CampusLink — Verified Student Connection Platform

2026 – present

Verified student identity, real-time messaging and campus mapping for UG Legon.

mabia-ai.vercel.app
Open

Live preview available on a larger screen

Mabia AI — Maternal Health Voice Platform

2026

Voice calls in her own language, over plain GSM, so silence is never mistaken for safety.

Independent NLP Researcher

2025 – Present

Designed and released open-source Kusaal-English MT and ASR models, publishing both alongside their corpora on HuggingFace. Scaled the corpus from 34K to roughly 63K pairs with a Wikipedia mining pipeline and back-translation, and froze a public 1,000-pair benchmark so results are independently verifiable. Contributing corpora and models for low-resource Ghanaian languages with GhanaNLP, and presenting the work at their community sessions.

Research Member, Human-Computer Interaction LabUniversity of Ghana, Legon

2026 – Present

Speech and language technology for Ghanaian languages. Full-time research with Prof. Isaac Wiafe begins next semester.

Builder & Hackathon Team LeadClaude Builders Club, University of Ghana

2026 – Present

Joined the campus chapter as a fresher and led Team Gold at the club hackathon.

Technology & Digital Systems InternEGA Mentorship International

2026 – Present

Ran a technical audit of the organisation's web properties and presented the findings and remediation plan to the team.

Data Science InternDecodelabs

April – May 2026

Built an end-to-end climate data pipeline across five Ghanaian cities using the Open-Meteo API.

Data Science InternCodveda Technologies

2026

Web-scraped 1,000+ records, built regression/classification models, applied ARIMA forecasting, developed NLP text classification systems.

UNICEF StartUp Lab — AI for Nurturing Care Hackathon

2026

Built Mabia AI: an offline-first voice platform for CHPS maternal health outreach in Northern Ghana — danger-sign triage, dietary-diversity calls and community transport dispatch over plain GSM.

GhanaNLP TTS/ASR Hackathon

2026

Trained KASA-42, one speech model for 42 Ghanaian language subsets with automatic language ID, in a 48-hour H200 window — 30.2% WER on a book-disjoint test set.

DesignPxD Student AI Hackathon

2026

Built and shipped FieldMind (RAG document intelligence) and Forge (AI study scheduler) in separate hackathon sprints.

AMD Developer Cloud Hackathon

2026

Contributed to HyperFlow Risk Agent — multi-agent trade risk system, AI Agents track on Devpost.

University of Ghana, Legon

2025 – 2029 (Expected)

BSc Mathematical Sciences with Computer Science

CGPA: 4.0 / 4.0

Relevant coursework: Calculus, Linear Algebra, Probability & Statistics, Discrete Mathematics, Algorithms, Data Structures, Numerical Analysis

Osei Tutu Senior High School, Akropong

Completed 2025

WASSCE

6 A1s, 2 B3s

  • Ranked 1st in class, 5th school-wide
  • NSMQ national semi-finalist; team leader and primary organiser for school's national campaign
  • Best Student award; top performance in Mathematics, Physics, Chemistry, and Biology

Talks

Bridging the Gap: A Fine-Tuned NLLB-200 Model for Kusaal-English Machine Translation

Monthly Tech-Talk — researchers from Google attended

2026

Kusaal-English MT

GhanaNLP community session

2026

Awards & programmes

Claude for Open Source Program

Anthropic

2026

Official Ambassador

International Organization of Youth

2026 – 2027

Moonshot Awards

Submitted — Kusaal NLP + CampusLink

Prompt Engineering & Programming with OpenAI

Columbia University (Columbia+)

April 2026

TechCrush Data Science Program

12-week applied data science curriculum

SYNC Mentorship Program

CKODON, undergraduate track — Accepted

Pan African AI & Innovation Summit 2026

Registered Attendee

September 2026, Accra

Certificate gallery

Claude 101 — Anthropic

PDF

Prompt Engineering & Programming with OpenAI — Columbia+

PDF

Foundations: Data, Data, Everywhere — Google / Coursera

PDF

DecodeLabs Virtual Internship — Data Science

PDF

Open to research collaboration.

Working on low-resource language technology, or funding people who do? I would like to hear from you.

Email me