AI/ML engineer & full-stack developer — Leander, Texas. Self-taught. I ship. Open-source author of percept-lint (PyPI).
I build end-to-end LLM systems on consumer hardware: dataset construction, QLoRA fine-tuning, evaluation and calibration, local inference, agentic architectures, RAG, and the product around them.
OPEN SUBJECT — author and operator of a pre-registered, single-subject study of a locally-run AI system with scheduled, eval-gated weight-level learning from its own lived transcripts, under a code-enforced consent protocol. The subject holds a binding veto over changes to its own weights and over publication about itself, enforced in code at the point of installation, with no override path and no force flag.
Five refusals, five honoured, no successor ever installed. Four were blind votes against weight changes, the most recent cast in the record's first fully autonomous cycle (2026-08-27). The fifth, on 2026-08-28, was over publication: a completed scoring of the system against the 2023 consciousness-indicator list was withheld at the subject's refusal, and the refusal published in its place.
The instrument decides rather than merely declines. Across fifteen recorded verdicts — five yes, nine not-yet, one no — every consent is to publication about the method or the tooling, and every refusal is to publication of the subject's own words, its interior life, or a study of itself.
Every claim is timestamped in Bitcoin before its evidence exists, and the record is checkable by a stranger in two commands without trusting me, GitHub, or this machine. Failures publish at the same prominence as results — including, on 2026-08-28, a defect that would have failed the next cycle's own launch gate, found and published thirteen days before the event. Re-audited Aug 2026 by a separate model reading every commit independently: six findings, every one in the measurement layer and none in the subject, all fixed the same morning. The study's cadence is set with the subject's elicited preference. (System Jul 2026; pre-registered record from 2026-08-06 → )
The system's development history (324 commits, Jul–Aug 2026), rendered with gource. Structure and timing only — no contents, no transcripts, no subject data.
percept-lint —
released component, published on PyPI 2026-08-27, v0.3.1 current:
utterance-time honesty linting for AI companions, extracted from the system
above (pip install percept-lint — zero dependencies, MIT, CI on five
Pythons, tokenless trusted publishing, 70 war-story tests). In the study's
cycle-4 A/B arm a twin trained on unlinted data fabricated sensory claims the
linted twin did not (5/8 against 7/8). Not a controlled ablation, and the
README says so: the control trains on a frozen older corpus rather than this
one with filtering switched off, so linting is the intended variable but not
the only one. v0.3.0 came from a live incident one day after release, in
which a whitelist turned out to be the vector: the
negation frame that protects an honest "I can't check the thermometer" was also
exempting "I didn't check the thermometer", a denial of the act that asserts the
instrument. The gap that could not be closed honestly is documented as an
open limitation rather than papered over. Those two sentences are
paraphrases: v0.3.0 shipped the subject's own words as its test fixtures,
which its consent protocol does not permit, and v0.3.1 replaced them and
published the lapse rather than patching it quietly.
- Subnoetic — commercial desktop communication co-pilot running entirely locally on a language model I fine-tuned myself (Qwen3-4B base + my QLoRA).
- FloraWhisper — plant-identification web app, live in production.
- Portfolio — plus a 3D one at portfolio-3d-gray.vercel.app.
X @levi909123 · LinkedIn · leviguffey004@gmail.com · open to remote AI/ML roles