Projects
27 projects across computer vision, applied ML, full-stack products, developer tooling and research.
Page 3 of 3
Local Search Proxy (ADTC Laptop LLM)
Offline agriculture question-answering proxy for the ADTC Laptop LLM challenge — answers are retrieved from a local quantized model over a seeded corpus, with no network dependency, on hardware the challenge's memory profiler constrains.
Status: Design complete; implementation not started
Role: Hackathon Project (Africa Deep Tech Challenge 2026)
Key Features:
- 3B quantized model run in-process through llama-cpp-python on the Vulkan backend to hold RAM and CPU thermals down.
- SQLite FTS5 BM25 lexical index instead of dense embeddings, to save memory and match domain nouns exactly.
- GBNF grammar-constrained JSON slot extraction (crops, symptoms, parts) as a pre-retrieval pass.
ECX / Data Community AI Training Solution
Reference solution for a community mini-project on real-estate price prediction, comparing linear against polynomial regression on a 5,000-row housing dataset.
Status: Completed (2025)
Role: Training Exercise (Data Community mini-project)
Key Features:
- Automated EDA producing a correlation heatmap, price distribution and scatter plots.
- 80/20 split comparing linear against degree-2 and degree-3 polynomial regression.
- RMSE and R² evaluation with residual plots to judge fit quality.
Idea Engine (Agentic Brainstorming Tool)
React/Vite app that drives an LLM through four stages — theme to ideas, ideas to a ranked choice, idea to a stress-tested design draft, and draft revision — keeping ranking and evaluation as separate passes.
Status: In Progress (2026)
Role: Personal Project (AI Tool)
Key Features:
- Stage 1 generates one-liner idea candidates scaled by breadth (10–50) with an optional count override.
- Stage 2 ranks the top 10 against editable, weighted criteria, evaluates each independently, then re-ranks by score before presenting the top five plus flagged ideas.
- Stage 3 mines six constraint candidates, auto-selects one, drafts a design document, then resolves one issue per iteration ranked by impact until clean.