The first three posts in this series were about how Sheria is built — the agentic pipeline, the semantic cache, a tour of the codebase. This one is different: less code, more plainly, what is Sheria actually for, and what does it do?
The problem, in plain terms
Kenya's courts have a documented backlog problem. Independent reporting has put pending cases at over 520,000 for stretches of the last few years, and even after a genuinely strong year — the 2024/25 financial year saw a 104% case clearance rate and a 27% drop in backlog, per the Chief Justice's own State of the Judiciary report — courts still resolved over 640,000 cases against more than 620,000 new filings in a single year. That's the scale of system this is built around: not a hypothetical, a live, currently-improving-but-still-enormous pipeline of cases.
Underneath the headline backlog number are more specific, everyday problems: judges and registrars spending large chunks of their time on manual legal research instead of writing judgments; court records still sitting in physical files or poorly scanned documents; and, less visibly, a real problem with document authenticity — forged court orders and fabricated evidence are a genuine risk when there's no fast way to check a document before relying on it.
I know the digitization half of this problem firsthand. My day job is Lead Data Engineer at ArdhiSasa, Kenya's national land management platform, where a large part of what I do is turning millions of physical land records into structured, searchable data. Sheria applies the same instinct to a different part of government.
The four modules
Sheria is organized around four modules, each aimed at one part of that pipeline.
Sheria Ask is the part most people would recognize as "the AI" — a chat interface where you can ask a legal research question in plain language and get back an answer grounded in actual Kenya Law Reports, with citations attached, rather than a confident-sounding guess. This is what the first two posts in this series were actually about, under the hood.
Sheria Digitize is the ingestion side: it takes court documents — PDFs, scans, plain text — and turns them into something searchable, chunked and indexed so a case doesn't just exist as an image nobody can search.
Sheria Verify checks whether a submitted court document holds up. It extracts metadata and runs a similarity check against known genuine documents, then returns a confidence score and risk flags rather than a bare yes or no. It's the module aimed most directly at the fraud problem.
Sheria Predict is the analytics piece: forecasting how long a case is likely to take and helping surface workload imbalances across judges. I'll be honest that this is the module furthest from finished — the API endpoints exist and are wired up, but the predictive model behind them is still being built out.
Where this leaves things
The last three posts explained how the Ask module works under the hood. This one is the plainer version of why any of that matters: a backlog measured in the hundreds of thousands, a digitization problem I already work on professionally in a different part of government, and a fraud problem that a confidence score can meaningfully shrink. The next posts will probably go back underground — Verify and Predict haven't gotten the deep-dive treatment yet.