Ask the agent what I build.

Mastery-tracking learning engine that spots where a learner is struggling and re-conditions its own prompts to the level they are actually at.

The education project inside Sudaverse: a localized LLM that teaches to the actual syllabus, in the dialect students speak, without sending their work off-device.

Local-first orchestration platform where specialised agents spawn for a task, do the work against open-source models, and dissolve afterwards.

Retrieval engine pairing chain-of-thought decomposition with a multi-agent retrieval loop — query rewriting, web search and vector recall under one planner.

Concurrent ephemeral agents and knowledge capsules that triage, classify and respond to alerts with the reasoning attached.

Local-first analytics on open-weight models: natural language in, generated SQL and interactive dashboards out, nothing leaving the network.

Cleaning, classification and normalisation pipeline that turns raw text into corpora fit for actually training on.

Streaming collection, processing and health monitoring across many sources, with the pipeline observability to trust what comes out.

Benchmarks tokenizer efficiency on Sudanese dialect text — the groundwork for treating a low-resource language properly.

Predictive flood intelligence for at-risk communities from public satellite data and open-source tooling.

CT-scan classification with MLflow and DVC wired through the whole pipeline, so the experiments are reproducible rather than anecdotal.