What we’re building—and the question connecting it all

Start with the hardware and a useful job

The R640 work is about making available hardware productive. That includes recovering a stable machine, establishing a usable local-model environment, and testing workflows small enough to understand completely. The local coding pilot connected retrieval, a proposed repair, protected tests, and review on a small Python project.

We are trying to build useful capability in layers. A working demonstration gives us something concrete to improve, while its limits tell us which test belongs next.

Choose a model by what it can actually do

The model projects explore memory, storage, latency, concurrency, and task quality. The pruning experiments produced smaller runtime artifacts; the concurrency comparison showed why batch throughput and individual responsiveness need separate measurements. Slow responses and timeouts stayed in the research record.

The evaluator repair made the central point clearer: an API request can succeed while the answer fails the task. The aim is to route a job to the smallest adequate tool or model using evidence about that job, and seek stronger assistance or human review when the evidence calls for it. That broader routing capability is a research goal, not a completed general-purpose system.

Make knowledge reusable without losing its source

The Wisdom Library indexing and private-memory design address a different constraint: useful research should remain useful after the conversation ends. Searchable sources, dates, provenance, and conflict handling can help the next task start from a better foundation.

The keyword index is a recorded milestone; the full vector phase was pending at that checkpoint. The small Riku vector index is a separate working demonstration. Keeping those boundaries clear is part of building a memory system that deserves confidence.

Turn the next idea into a test

The 69-question AI catalog covers model efficiency, CPU systems, context, retrieval, verified tools, adaptation, and research integrity. It is an agenda with explicit next steps, rather than a list of completed breakthroughs.

Our objective is dependable, resource-aware AI that can do useful work, explain what supports its answer, preserve what went wrong, and stop when a result has not been established. I direct that work with AI-assisted implementation and review. The blog documents both the projects and the learning that connects them.

Sources & reading notes

1. Riku R640 local AI pilot — technical guideThe bounded local coding and retrieval demonstration. Drive record · access may be restricted Source record retained privately.
2. LLM Benchmark Study MasterTask-level model evaluation and evidence context. Drive workbook · access may be restricted Source record retained privately.
3. R640 Autonomous Operating Plan v1The research agenda and proposed operating model. 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.