Research with
a record behind it.

JIM BUTCHER, PhD MBA
Independent researcher & software developer
R640 & AI Research | Updated October 15, 2026

Research direction by Jim Butcher, with AI-assisted implementation and review.

PROFILE
Mission: Advance dependable, accessible AI through practical experiments, thoughtful engineering, and evidence we can stand behind.

Practical research into local AI, CPU/system-memory inference, retrieval, controlled code changes, and reproducible evaluation.

TOOLS AND METHODS
Python, Node.js, Linux, SQLite FTS, Qdrant, llama.cpp, LM Studio, GGUF, unittest, Benchmark design, Source provenance, Research protocols

SELECTED EVIDENCE
A local coding assistant with evidence and protected tests (2026-09-30)
A small R640 demonstration joined local retrieval, a generated Python fix, protected tests, and a reviewed apply step into one traceable workflow.
31 — Package regression tests passed; 8 / 8 — Demo tests after the reviewed fix; 12.69 tok/s — Reported generation rate · one small job
Evidence setting: R640 · historical pilot described in the September 30 guide, archived in Drive October 10.
https://jimbutcherstudio.com/r640/posts/riku-r640-pilot/

A successful request is not a correct answer (2026-10-07)
Seven successful requests did not mean seven correct answers. Repairing the evaluator changed what the benchmark could honestly claim.
31 / 31 — Evaluator regression cases passed; 7 — Successful requests replayed; 2 fail · 5 review — Answer-level reassessment
Evidence setting: Windows source PC; no inference or generated code executed in reassessment
https://jimbutcherstudio.com/r640/posts/quality-evaluator-repair/

Pruning LFM2: smaller weights, measured limits (2026-09-11)
Reducing an MoE model’s expert count produced a smaller runtime artifact and a modest latency difference in two small historical runs.
64 → 56 — Experts retained; 1.61 GiB — Storage saved; 4.495 → 4.006 s — Mean job time · small source-PC runs
Evidence setting: Windows source PC via LM Studio localhost API; not R640
https://jimbutcherstudio.com/r640/posts/lfm-expert-prune/

Making a large library searchable, one evidence layer at a time (2026-09-28)
A substantial keyword index is useful before a vector phase is finished. The Wisdom Library records make that distinction explicit.
13,468 — Gutenberg books in the recorded FTS index; 1,883,249 — Indexed chunks; 0 — Embedded chunks at that checkpoint
Evidence setting: HP Omni source-side indexing and Drive staging for R640 · September 28 records.
https://jimbutcherstudio.com/r640/posts/wisdom-retrieval-index/

Sixty-nine questions for dependable local AI (2026-10-14)
The research catalog covers model efficiency, CPU systems, memory, tools, and evaluation. A question belongs in the portfolio before it earns a result.
69 — Research questions catalogued; 64 — Draft research designs; 4 — Engineering-preparation records
Evidence setting: October 14 R640 operating-plan proposal, based on the recorded 69-owner portfolio.
https://jimbutcherstudio.com/r640/posts/sixty-nine-research-questions/

RESEARCH AGENDA
69 documented questions; 64 draft designs, four engineering preparations, one blocked preparation as of October 14. No accepted scientific results claimed for these 69 cards.

WORLD QUESTIONS
10 documented applied projects and7proposed future studies: food loss, household energy and spending, PFAS models, disability travel, regional industry, materials, finite mathematics, and symbolic communication. Recorded stages and limits: https://jimbutcherstudio.com/r640/world-questions/

CONTACT
https://www.linkedin.com/in/jim-butcher-phd-mba-0208b621a/

Full portfolio: https://jimbutcherstudio.com/r640/
Source records are retained privately; their links and access details are not published here. The portfolio separates historical observations, measured small pilots, research reviews and future designs. It is not a live status report or a full employment/education history.