/ notes

work

Work

Applied projects, research, and experiments across AI systems: agents, conversational and voice interfaces, and AI in education.

applied work

graspy
Multi-agent, multilingual AI tutor for students to learn in the language they understand best, with a Nigerian-languages variant (natlas). graspy.tosinamuda.com · github
voice agent
End-to-end voice interaction under 4s speech-to-speech latency. ibm.biz/innovation-studio-assistant
design thinking generator
Production generator for design-thinking artefacts; extended from CHAI-DT. live demo
karia
LLM-powered career exploration for students. ~2,000 users. karia-ai.web.app · github
llama slide generator
DSPy GEPA pipeline for synthesising consulting one-pager training data, then fine-tuning Llama 3 on AWS SageMaker. github

msc research

Design and Evaluation of a Lesson-Plan-Driven Framework for Curriculum-Constrained Personalised Learning

Ongoing · expected June 2026 · University of Lagos

Whether teacher-authored lesson plans can constrain LLM generation without losing curriculum alignment. Case study in Nigerian JSS1 mathematics, Design Science Research methodology.

studies & experiments

self-evolving agent skills
Whether agents can learn reusable procedural corrections from grounded repair conversations and use them to improve first-response performance on future same-family tasks.
params.ai
Comparing chat-style copy-paste against {{variable}} templates that auto-generate input forms. Tests whether structured forms outperform chat for reusing AI prompts. github
memolet replication
Replication of Yen & Zhao's Memolet (UIST 2024): long-term memory repository, semantic clustering, and prompt composition for reusable AI conversational memories. github
rlm playground
Full-stack playground for exploring Recursive Language Models against the LongBench dataset, with a UI to inspect RLM trajectories. github
baguettotron interface
Open-source local-CPU chat interface for Pleias' Baguettotron and Monad models. Custom parser renders the special thinking tokens these models emit as collapsible reasoning traces. github