FRANK DUAH2026NORTHEASTERN

--:--

Engineerbuildingintelligentsystemsthatship.

Full-stack and ML. Currently shipping reinforcement learning research and building AI agents.

01Index

02Selected work

2026Co-Author

Latency-Aware Execution Engine

Reinforcement learning trade execution

End-to-end ML system combining classical methods (Linear Regression, Random Forest, XGBoost) with QR-DQN reinforcement learning agents for optimal cryptocurrency trade execution. Beats TWAP by 10.2 basis points on Bitcoin walk-forward and cuts cost 96.7% versus naive execution, validated against 98 million real market trades across three assets. Implements the Almgren-Chriss market impact model and a risk-sensitive RL framework, with 17 ablation studies and 176 unit tests. Dual final project for CS5130 (Applied Programming) and CS6140 (Machine Learning) at Northeastern.

vs TWAP (BTC): +10.2 bps
Cost reduction: 96.7%
Validation: 98M trades
Tests / Ablations: 176 / 17

Python
PyTorch
QR-DQN
XGBoost

2025Co-Author

Music Emotion Recognition

Transfer learning with RoBERTa for valence and arousal

Transfer learning with a 125M parameter RoBERTa transformer to predict valence and arousal dimensions from song lyrics. Custom 4-layer neural network with dropout and early stopping. Reaches 0.781 valence correlation and 0.649 arousal correlation on the 2,568-song MERGE dataset, performing at the top end of published research benchmarks. Also generates 128-dimensional embeddings for downstream multimodal ensemble learning that combines lyrics with audio features.

Valence correlation: 0.781
Arousal correlation: 0.649
Backbone: RoBERTa 125M
Dataset: MERGE 2,568

Python
PyTorch
RoBERTa
Transfer Learning

2026Contributor

llm-exe

Open source contributor

Major contributor to the open source TypeScript framework for production LLM applications. 30+ merged pull requests spanning release pipeline hardening, workflow automation, and documentation infrastructure. Built the AI agents that now help self-maintain the repository: automated reviewers, docs-sync workflows, and release-prep bots that keep the framework healthy with minimal human intervention.

Merged PRs: 30+
Specialty: Self-maintaining agents
Areas: Agents / CI / Docs

TypeScript
GitHub Actions
LLM Agents

2024Contributor

Brilla AI

Open source AI that competes live in Ghana's National Science and Maths Quiz

Contributing engineer on the open source AI that competes live alongside students in Ghana's National Science and Maths Quiz — combining speech-to-text, question-answering, and text-to-speech behind a real-time webapp. Built and shipped video player, layout, and frontend control updates for the team's broadcast interface.

Stars: 34
Modalities: STT · QA · TTS
Surface: Live broadcast webapp

TypeScript
React
Python
Jupyter

2022Co-Author

Spark

Safer social platform for kids, with ML age verification

Final-year thesis at the University of Ghana. Built a mobile social platform deliberately stripped of likes, comments, and followers, with real-time messaging, audio and video calling, and media sharing in Flutter and Firebase. Worked alongside teammates on the ML age verification pipeline: MTCNN face detection, Inception v2 plus FaceNet face verification, easyOCR for ID parsing, and VGG16 transfer-learning classifiers at two age thresholds. Supervised by Dr. Nii Longdon Sowah.

Surface: Mobile (Flutter)
ML pipeline: MTCNN + FaceNet + VGG16
Backend: Firebase
Thesis year: 2022

Flutter
Dart
Firebase
Python
TensorFlow
OpenCV

2026Author

AI-PR-Assistant

GitHub-integrated review bot

A bot that summarizes pull requests, highlights important changes, and suggests improvements. Wired directly into the GitHub workflow so reviews happen automatically alongside human ones.

Surface: GitHub
Integration: Native PR

TypeScript
GitHub Actions

03About

I build systems that move from idea to running production code. Most of my work sits at the intersection of full-stack engineering and applied machine learning: reinforcement learning agents for trade execution, transformer-based music emotion recognition, AI agents that help maintain the repositories they live in, and quiet contributions to the open source LLM tooling other developers depend on.

Pursuing my MS in Artificial Intelligence at Northeastern University on a fully-funded merit scholarship for which I was the sole graduate of the University of Ghana selected. Before that, a First-Class Bachelor's in Computer Engineering and the Dean's Award four years running (2019–2022). I treat "production-ready on first pass" as a rule rather than an aspiration.

Most of the work above shipped with teams: research partners on transformer modeling, a multi-disciplinary squad behind Ghana's National Science and Maths Quiz AI, a thesis trio for the social platform. I lead when I should and follow when I should.

0+
years writing code
0+
repos
0+
merged OSS PRs
location

42.3398° N, 71.0892° W

04Process

How I work.

01ResearchPapers, issues, tickets, real users.
02PrototypeSmallest thing that proves it works.
03MeasureNumbers before opinions.
04ShipProduction-ready on the first pass.
05IterateEvery system gets better in v2.

05Education

2025 — present

Northeastern University

Master of Science in Artificial Intelligence

Sole graduate of the University of Ghana selected for a fully-funded merit scholarship covering tuition and living expenses.

2022

University of Ghana

Bachelor of Science in Computer Engineering

First-Class. Dean's Award for Best Performing Students three consecutive years (2019, 2020, 2021).

06Writing

May 2, 2025 · 9 min · LLM · TypeScript · Open Source

Supercharging LLM Applications with llm-exe

A walk through the llm-exe SDK with three concrete patterns: validating statements against conversation context, extracting structured data from unstructured text, and detecting user intent. One toolkit, four providers, modular components.

Read on Medium

07Today

May 24, 2026

Data Engineering

Today, I taught myself

data versioning with DVC

DVC (Data Version Control) is Git's cousin for data and ML models. Instead of storing 50GB datasets in your repo (which Git hates), DVC stores a tiny pointer file in Git and keeps the actual data in cloud storage like S3 or GCS. Think of it like GitHub storing a reference to a Docker image instead of the whole image—Git tracks the `.dvc` file, DVC tracks what's really there. It also tracks pipelines: the steps to transform raw data into features for your model. So if your training data changes, you can see what broke and retrain, not just wonder why last month's model doesn't work anymore.

Git for data: version your datasets like you version code, track your data pipeline like you track code changes.

Lock it in

Why can't you just store your 10GB training dataset directly in a Git repo alongside your model code?

08Contact

Tell me what you are building.

Hiring, collaboration, or a question about the work above. I read everything.

Reinforcement learning that beats production algos.

Engineerbuildingintelligentsystemsthatship.

01 / Index

02 / Selected work

03 / About

04 / Process

How I work.

05 / Education

Northeastern University

University of Ghana

06 / Writing

Supercharging LLM Applications with llm-exe

07 / Today

data versioning with DVC

08 / Contact

Tell me what you are building.

07 / Today

data versioning with DVC

01Index

02Selected work

03About

04Process

05Education

06Writing

07Today

08Contact

07Today