A local coding assistant with evidence and protected tests

Package regression tests passed

31

Demo tests after the reviewed fix

8 / 8

Reported generation rate · one small job

12.69 tok/s

The working demonstration

The Riku pilot used an existing Qwen3-Coder 30B Q4 model, a separate Qwen3 embedding model, and a dedicated Qdrant database. It indexed a small calculator project, retrieved relevant source, proposed a one-line calculation repair, and applied the change after review and protected testing.

The saved guide reports that eight baseline tests ran, seven failed on the deliberately introduced bug, and all eight passed after the reviewed change. Test-file hashes stayed unchanged.

What was measured

The code-generation job took 27.984 seconds at a reported 12.69 generated tokens per second. Initial indexing covered four files and three chunks in 3.763 seconds. The correct source ranked first in a retrieval taking 0.426 seconds, including query embedding.

After the change, incremental indexing re-embedded one changed file in 0.555 seconds. The source-based answer was checked against the calculator’s formula and tax behavior.

Why the boundaries matter

The pilot separated proposals from accepted source changes, preserved the test suite, and checked source citations. Its guide also records live limits and authenticated local endpoints at the time of the demonstration.

Cold reboot recovery, sustained performance, and use on a real repository remained unvalidated. This historical guide establishes a small demonstration, not today’s service state or a general coding-success rate.

An additional research workflow

The same guide records a first web-research task covering ten Houston oilfield businesses. The R640 collected 17 official public pages, drafted records locally in two batches, and checked the draft fields against the sources before export. This is a bounded example of evidence-led research, not a general web-accuracy benchmark.

What we are trying to achieve

The longer-term aim is a local assistant that can understand an approved project, retrieve the right evidence, propose a small change, and demonstrate why that change is acceptable. The calculator pilot connects those parts on a deliberately small example. Broader repositories, recovery, and sustained operation need their own evidence.

Sources & reading notes

1. Riku R640 local AI pilot — technical guideVerified results; first web research job; limits. Drive record · access may be restricted Source record retained privately.

This is a dated, source-based research account, not a live machine report. Historical results, engineering preparation, research designs, open review gates, and measured pilots remain distinct.