RRLab is Rodney Rui's independent research lab.
Directions: multi-model collaboration, AI evaluation & selection,
AI-native products.
Research findings become usable tools;
tools power engineering delivery.
A wheeled life-companion concept: follows you, can lead the way, voice commands on the roadmap. Lifestyle form. Embodied AI within reach.
A single model has limits. RRLab does not search for a stronger
model — it asks how to organize multiple models, each with its
own strengths, into a system that delivers reliably within known
boundaries, produces verifiable new knowledge beyond them, and
generates its own governance rules.
Four research lines, each with papers and runnable public evidence.
Explicitly decompose what capability a task needs, then route
directly to the strongest model along that dimension, without
intermediate representations. Reasoning depth improves by 25.9%
over implicit routing.
A purely classical framework inspired by the measurement
structure of quantum mechanics: before a decision, the system
retains a distribution over plausible worlds, instead of
collapsing prematurely into a single confidence score. That
distribution maps to an auditable route: auto-accept, review,
fallback, abstain. The goal is not to prevent AI from being
wrong — it is to prevent AI from being wrong with high confidence.
Hallucination cannot be dammed, but it can be channeled. In
exploration mode, multiple models independently analyze the
same object from complementary perspectives — not seeking
consensus, but detecting resonance: structurally compatible
interpretations reached from different paths. Resonant
directions enter a computational falsification queue. In the
BSD campaign, 6/7 strongly resonant hypotheses survived
subsequent verification.
Orchestration rules are not stacked from human presets — a
reasoning model derives them deductively from system invariants
alone. Ablation: deductive 96/100, inductive 76/100. In a
larger experiment, a system given only minimal constraints
(explore + detect resonance) self-organized dependency-aware
batching and cross-model file sharing — the paper calls it
"ordered behavior under minimal constraints": observed, not
explained. Consistent cross-project experience: adding human
micro-constraints degraded output quality.
The runtime that carries the four research lines. Multi-model
orchestration with a multi-dimensional evaluation layer
(50-task benchmark), supporting both deterministic and
exploratory modes. Largest experiment to date: 15 skills,
7+ heterogeneous models, 77 tool calls, 74 minutes of
autonomous operation.
Full-stack productization of multi-model orchestration.
eVoiceClaw (iOS voice assistant, multi-provider switching),
Desktop (multi-agent workstation), Pro (fully local knowledge
base for doctors and engineers), Cerebellum (on-device
"cerebellum" model: semantic routing + privacy awareness +
safety review). Built end-to-end: Swift frontend, Python
backend, agent orchestration.
Agent code-modification capability audit. It does not rank
models; it answers two engineering questions: after a code
change, do previously passing tests still pass (FRR)? At equal
quality, who is faster and cheaper?
Multi-dimensional LLM benchmark for Chinese AI scenarios.
89 questions × 13 dimensions, 4 difficulty tiers (incl. expert)
with weighted scoring, 6 scoring methods, 12 providers preconfigured.
Complements RRLab-Bench: general capability baseline, agent-grade engineering.
Uncertainty routing engine: decomposes landscape uncertainty
into two orthogonal components — occupancy (is there signal
at this scale?) × amplitude (how strong?). Detects toward-1
hardening (high-confidence false reassurance), fixing the
coverage collapse of classical BPL bands. Domain-neutral.
Reference application of the BPL/OA engine: stock-screening QA.
Does not predict stocks — audits the screener's output
distribution: batch health, per-stock confidence flags,
threshold calibration. Method and code fully public.
Layer 1 — Mac-TTS (v0.1 released, open source): turns every
Mac's built-in speech into a Unix pipe. Zero disk I/O, pure
CPU inference, TTS + STT in one. Layer 2 (in development):
emotion-controlled multi-character voice acting, currently
via cloud TTS.
From "benchmarking models" to "benchmarking and evolving the
harness": how the runtime around a model — workflows,
evaluation, permissions — can itself be systematically
evaluated and improved.
Digital asset & will management app. Asset registration,
beneficiary management, liveness confirmation & delivery,
verification-code identity checks.
Free / annual / VIP tiers. iOS app & backend API in parallel.
Original-novel listening app for short-drama fans.
Import the source TXT after finishing a drama;
chapter navigation + tiered TTS.
Endgame: per-character AI voice casting — role detection,
voice assignment, sentence-level synthesis.
AI industry deep-reading intelligence station. Collects 100+
technical blogs, arXiv papers, HackerNews threads and GitHub
repos daily, scored by DeepSeek V4 Flash on 3 dimensions,
selected per-category with separate quotas, and digested into
structured reading notes. Fully automated, runs at 3:00 AM daily.
Global news aggregation dashboard. China hot lists (Weibo,
Zhihu, Douyin and 14 more platforms) + GDELT global news
+ HackerNews + RSS + financial news + A-share earnings.
SimHash smart dedup, 10-min auto refresh, React responsive
dashboard, one-click Docker deployment.
On-device semantic routing model (self-trained). BGE-M3 base
+ 15-dim intent regression head, 51.4M trainable params,
~1GB resident — turns user messages into intent vectors,
replacing LLM calls in agent routing. Millisecond latency.
Code open-sourced + 202MB increment on HuggingFace.
Wheeled life-companion concept: follow, lead, voice-command roadmap; tagline Follow Your Life.
Concept views, follow film, and full theme song are live.
Vertical doorplate-address recognition. YOLOv8 detection
+ PaddleOCR recognition + Chinese-number normalization
(「二十」→「20」) + address assembly. Self-built auto-labeling
pipeline cuts manual labeling by 90%, data scaled to 100k+;
accuracy acc 0.9962. Delivered as an FDE consulting case.
Empirical study of causal chain between model identity misattribution
and task failure, using bench-harness discriminative probes. Lab-reproduced:
3/4 runs, model reads PI_MODEL=deepseek yet outputs contain "Claude".
Contextualized within Spartacus (25.93% confusion rate) and ICML 2026
Role Confusion (position decay effect). Full data loop established:
observation → literature verification → reproducible probe → fix strategy.
Preliminary finding: no architecture-level cure exists, but the combination
of developer-message identity fields + runtime enforcement + periodic
re-injection can approach a practical resolution.
All paper preprints are on Zenodo; every DOI is verifiable.
ORCID: 0009-0004-3898-4655
validated on the self-built multi-model orchestration system
application of the resonance-exploration mode; large-scale SageMath computation + Magma verification + AI-assisted research
Independent Forward Deployed Engineer.
RRLab's research is not paper theory — the four orchestration
modes, benchmark evaluation, and uncertainty routing are
engineering methods that can be embedded directly into your
systems. I work embedded in your team and take LLMs from demo
to production-line productivity.
Mathematician × AI systems builder × Independent FDE.
Academic background in computational number theory: numerical verification methods for elliptic curves and the BSD conjecture. Recent focus: AI systems engineering — multi-model orchestration, agent evaluation, uncertainty routing. RRLab is the unified vehicle for this work.
This methodology was not read from reports. In the BSD campaign, a multi-model AI system produced publishable discoveries — resonance proposed hypotheses, deterministic computation falsified them, and 13 papers were the byproduct. Built the eVoiceClaw product family from scratch across four product lines. Self-built benchmarks quantify model capabilities; every selection decision has a measurement behind it. Trained an on-device intent-routing model (Cerebellum, on HuggingFace), turned the uncertainty-routing method into a reusable engine (BPL/OA, open source), and delivered vertical projects such as doorplate OCR. Research conclusions become tools, and the tools are open source — 9 repositories in total.
Frontier exploration on genuinely open problems, with results — most business problems are easier than this.
Have an AI idea and don't know where to start? Let's talk — no pressure.