FRANK DUAH2026NORTHEASTERN
--:--
I hand-build the full LLM lifecycle: a transformer, a tokenizer, fine-tuning, an MCP server, and an agent red-teaming harness, all with no frameworks, then pair that depth with real software engineering.
I build the LLM stack from the ground up to understand it, then ship production systems on top of that understanding. A transformer with every gradient derived by hand, a byte-level tokenizer, LoRA and its modern successors, an MCP server, and a red-teaming harness that fails CI when an agent gets less safe. No frameworks in any of them, and the code is public and readable.
That depth sits next to ordinary software engineering: full-stack work, applied machine learning, and open source. I am the second-highest contributor to llm-exe, a production LLM framework other developers build on. I iterate hard, and I hold every iteration to a production bar so the next pass builds on solid ground instead of a workaround.
Pursuing my MS in Artificial Intelligence at Northeastern University on a fully-funded merit scholarship, the sole graduate of the University of Ghana selected for it. Before that, a First-Class Bachelor's in Computer Engineering and the Dean's Award four years running (2019 to 2022). Much of the work is collaborative: research partners on transformer modeling, the multi-disciplinary team behind Ghana's National Science and Maths Quiz AI, and a thesis team that engineered a social platform from the ground up.
Skills
7+
years writing code
40+
location
The full LLM lifecycle, implemented by hand with no frameworks: a transformer, efficient fine-tuning, a tokenizer, and a Model Context Protocol server.
01
2026Author
A transformer from scratch, every gradient derived by hand
A working transformer implemented in NumPy with no autograd anywhere. It supports both GPT-2 and Llama-style architectures (RMSNorm, RoPE, SwiGLU, and a KV-cache), and every forward and backward pass is hand-derived and gradient-checked to 1e-8. It trains and generates text on a plain CPU.
02
2026Author
Modern fine-tuning methods rebuilt from the math up
LoRA, DoRA (the 2024 method), and QLoRA-style int8 quantization, all implemented by hand. The forward and backward passes are hand-written and gradient-checked, adapters hot-swap at runtime, perplexity evaluation is built in, and the frozen base model is provably left untouched.
03
2026Author
A byte-level BPE tokenizer you can watch work in the browser
A byte-level Byte-Pair Encoding tokenizer built from the standard library, with a provably lossless round-trip on arbitrary Unicode. It exports a dependency-free in-browser tokenizer that matches the Python encoder token for token.
04
2026Author
A Model Context Protocol server with no SDK, plus an injection scanner
An MCP server implementing JSON-RPC 2.0 over stdio with no SDK. It compacts documents to Markdown (25 to 75% fewer tokens) and scans them for prompt injection before they reach a model, and it connects to Claude as a tool.
05
2026Author
An adversarial stress-tester that fails CI when an agent gets less safe
An adversarial red-teaming harness for LLM agents, built from scratch with zero runtime dependencies. It runs hostile and edge-case attacks aligned with the OWASP LLM Top 10, reports per-category hold rates with the exact inputs that break the agent, and gates CI so a build fails when a prompt or model change makes an agent measurably less safe. Hold rates run over N trials, inconclusive judge calls are excluded from the denominator, and heuristic detectors are labeled as such.
06
2026Author
Offline-first retrieval that picks its own model backend
A retrieval-augmented assistant whose defining idea is that the SAME code detects its runtime environment - network plus CPU and RAM - and automatically chooses a backend: a cloud model when resources are free, a quantized on-device model when they are not, with no user-facing toggle. Hybrid retrieval narrows by date before semantic search, and a pinned, model-aware embedding cache keeps vectors from different models from ever mixing. The live demo on this site is a browser-only slice of the same idea: it embeds every post on your device with a quantized MiniLM model and ranks by cosine locally, so ranking costs nothing per query and needs no search backend.
07
2026Co-Author
Reinforcement learning trade execution
End-to-end ML system combining classical methods (Linear Regression, Random Forest, XGBoost) with a Double DQN reinforcement-learning agent (dueling architecture, prioritized experience replay) for optimal cryptocurrency trade execution. Beats TWAP by 10.2 basis points on Bitcoin walk-forward and cuts cost 96.7% versus immediate execution (the naive baseline), validated against 98 million real market trades across three assets. Implements the Almgren-Chriss market impact model, with 17 ablation studies and 176 unit tests. Dual final project for CS5130 (Applied Programming) and CS6140 (Machine Learning) at Northeastern.
08
2026Collaborator
Multimodal embedding input, merged upstream
Second-highest contributor and a Collaborator on llm-exe, an open-source TypeScript framework for building LLM applications. Authored its multimodal embedding input feature: one provider-neutral input type routing to AWS Bedrock Cohere Embed v4, while text-only providers reject image content at the boundary with a typed error. Merged upstream and published in v3.0.3, additive and non-breaking.
More work
Transfer learning with RoBERTa for valence and arousal
0.781 valence and 0.649 arousal correlation on the MERGE dataset's 2,568-lyric split.
Safer social platform for kids, with ML age verification
A child-safe social platform with a multi-stage ML age-verification pipeline (MTCNN, FaceNet, VGG16), built as a 3-person BSc thesis.
Open source AI that competes live in Ghana's National Science and Maths Quiz
A live AI that competes in Ghana's National Science and Maths Quiz; I contributed the frontend and AI integration.
GitHub-integrated review bot
Summarizes pull requests and flags risky changes inline, wired into the GitHub review flow.
01
2025 - present
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.
02
2022
Bachelor of Science in Computer Engineering
First-Class. Dean's Award for Best Performing Students four consecutive years (2019, 2020, 2021, 2022).
May 2, 2025 · 9 min · LLM · TypeScript · Open Source
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 MediumMost days I post what I am learning about AI and agents. Follow along on LinkedIn, or read the archive here.
Hiring, collaboration, or a question about the work above. I read everything.
Agentic AI
Today, I taught myself
An "agent orchestration framework" is just a library that manages the traffic control between multiple LLM calls — deciding what runs next, what state gets passed along, and when a human or another agent needs to step in. Think of it like the difference between air traffic control (strict rules, every plane's position tracked) and a group project where everyone just knows their role and improvises (loose, fast to set up). LangGraph (from the LangChain team) is the air-traffic-control option: you define your agent as a graph of "nodes" (functions that do work) and "edges" (rules for what happens next), and it tracks a shared "state" object as data flows through — giving you fine-grained control, retries, and pause/resume for human approval. CrewAI is the group-project option: you describe agents by role, goal, and backstory (like hiring for a team — "you're the researcher, you're the writer"), assign them tasks, and the framework figures out the coordination. OpenAI's Agents SDK sits in between: a thin, minimal-ceremony toolkit built around two ideas — agents and "handoffs" (one agent passing control to another) — that grew out of an earlier educational project called Swarm (released 2024, now retired) and became OpenAI's production-recommended framework in 2025.
LangGraph = you write the flowchart. CrewAI = you write the job descriptions. OpenAI Agents SDK = you write the minimum.
Lock it in
You're building a multi-step agent system where a human needs to approve a risky action mid-workflow before it continues. Which framework's model — LangGraph's graph-with-state, or CrewAI's role-based crew — more naturally supports that kind of "pause and wait" checkpoint?
7200 contributions · 6299 in private repos