Cases Method Services Contact Home Blog Digest 中文
FDE — FORWARD DEPLOYED ENGINEER · INDEPENDENT

Taking AI from
demo to production.

Independent FDE. One person, no middle layer.
Delivering runnable, quantifiable, maintainable AI systems —
methodology all from public research, results all benchmark-verifiable.

Email · 30-min alignment Evidence AIGC · Resonance AIGC · Pet Friends R2DF
202/202
All subtasks passed in a 50-task agentic benchmark
96 : 76
Ablation: deductive rule generation > inductive learning
25.9%
Reasoning depth gain from dimension-direct routing
6/7
Strong-resonance hypotheses surviving verification (BSD)
1,682
Tool calls in a single research campaign (43 runs / 302 skills)
13
Zenodo preprints, methodology all public
01

Cases
Evidence, not promises

01

Mathematical Discovery with a Multi-Model AI System

Domain: Computational Number Theory (BSD Conjecture) · Role: System Designer & Conductor

Problem

The BSD Conjecture is a Millennium Prize problem. On the massive Cremona database of elliptic curves, we needed large-scale computation, pattern discovery, and hypothesis falsification — no single model could handle this alone.

Approach

43 runs, 302 skills, 1,682 tool calls. Exploration mode: multi-model divergence → resonance → 13 annotated hypotheses. Deterministic computation handled falsification. 6/7 strongly resonant hypotheses survived verification. One decisive fact discovered deterministically → exploration mode generated explanations → three papers in 43 hours.

Verifiable Results

8 computational number theory papers on Zenodo; a complete methodological longitudinal case study (Part 4), including all fabrications by the AI drafting layer — published, not hidden.

What This Proves

On problems where the answer does not exist in any training data, this system produces new knowledge that survives verification. Your business problem is almost certainly easier than the BSD Conjecture.

02

eVoiceClaw: Full-Stack Delivery of Four Product Lines

Domain: Consumer AI Products · Role: Sole Developer

Deliverables

eVoiceClaw — iOS 18 voice assistant (Swift 5.9), multi-provider switching (Claude / OpenAI / Gemini / custom CrewAI backend), multi-agent coordination. Desktop — multi-agent collaborative workstation, production implementation of dimension-direct routing. Pro — fully local knowledge base (local models + vector search + local TTS) for data-sensitive professionals. Cerebellum — on-device "cerebellum" model: semantic routing, privacy awareness, safety review — data never leaves the device.

What This Proves

Not just a paper-writing researcher. From Swift frontend to Python backend to agent orchestration to on-device models — every layer, solo-delivered. There is no layer in your project I haven't touched.

03

Replacing Gut Feelings with Benchmarks

Domain: AI System Evaluation · Role: Evaluation Framework Designer

Approach

Self-built 50-task multi-dimensional agent evaluation dataset covering code construction, information retrieval, multi-step reasoning. Supports multi-model cross-comparison, routing strategy A/B testing, and independent evaluation of both modes (deterministic / exploratory).

Verifiable Results

Performance differences between orchestration strategies are quantifiable — deductive rule generation 96/100 vs. inductive rule learning 76/100. "The right model for the right task" isn't a slogan — it's a data-backed decision every time.

What This Proves

Every selection recommendation I make comes with a reproducible measurement method — not "in my experience."

04

Doorplate OCR: Vertical Solution from Detection to Recognition

Domain: Computer Vision · Role: Full-Stack Developer (Data, Training, Deployment)

Problem

Doorplate address recognition: complex Chinese address formats (road/alley/branch/number), tilted shots, uneven lighting, blurred plates — generic OCR falls short. Needed a complete vertical solution: detection + recognition + address structuring.

Approach

YOLOv8 plate detection → crop → PaddleOCR text recognition → Chinese numeral normalization ("二十" → "20") → address assembly. Self-built auto-labeling pipeline (YOLO crop + PPOCRLabel) cut manual labeling by 90%, scaling from thousands to 100k+ samples. Training used cloud computing + large-model knowledge distillation.

Verifiable Results

Recognition accuracy: 0.9962, normalized edit distance: 0.9990 (at 5,000+ training samples); further improvements with 100k+ data. End-to-end delivery: labeling → training → accuracy达标 → deployment.

What This Proves

Complete vertical OCR delivery capability: data, training, accuracy optimization, deployment. Doorplates worked — your receipts, labels, IDs: same class of problem, same class of method.

05

Legacy Access Control: Offline Issuing After Vendor Sunset

Domain: Legacy Systems Engineering · Role: Reverse Engineering + Tooling (CLI / GUI)

Problem

A residential site issued cards through a vendor smart-access platform. The vendor service term had expired and the vendor no longer supported the product. The client did not want a full hardware/system replacement in the near term—but after the cloud issuing path failed, existing cards still opened doors while new cards could not be issued.

Approach

Under client authorization, preserved the original workstation software and multi-year logs; reverse-engineered the write path and showed card encryption and door-side checks do not depend on the cloud—so fully offline issuing is feasible. Bulk log decryption recovered the permission model (resident vs. admin cards). Card air interface identified as ISO15693 (not the more common 14443A/Mifare white-cards). Delivered a Windows offline issuer (self-tests + byte-identical golden frames vs. the original write component) plus a GUI matching prior operator habits, and migrated a searchable historical issue archive.

Verifiable Results

Crypto round-trip self-test passed; resident and admin frames match the original write component golden standard byte-for-byte. Thousands of unique frames parsed from logs; 5,000+ records migrated into a local archive. Deliverables: CLI, Chinese GUI, docs. Door-side closure (write a live blank → open a door) awaits on-site encoder and blank-card acceptance; external narrative will be strengthened after that step.

What This Proves

FDE work is not only greenfield AI products. When a vendor sunsets support and the client will not rip-and-replace, the job is to turn “still opens, cannot issue” into handoff-ready local tooling—with golden tests and log-scale evidence, not another opaque black box.

06

R2DF by RRLab: Follow Your Life · Life Companion on Wheels (Concept)

Domain: Personal robotics / life assistant · Tagline: Follow Your Life

Problem

People need more than another set of wheels—they need a small life companion: carry, lead the way when you want a guide ahead, and eventually take voice commands. Senior-care styling turns buyers off; lab robots stay out of reach.

Approach

Tagline: Follow Your Life. Vision: a reliable wheeled sidekick (the helpful-companion spirit of classic service droids—without the movie IP). Capability stack: follow-and-carry → lead (cart ahead, you follow) → voice commands (roadmap). Form: covered box, sit-or-carry, lifestyle not medical. Bounds: low speed, always cancelable; no stairs, no seated rolling, no unsupervised city roaming. Concept views and film are live.

Verifiable Results

Public concept page with front / side / rear stills and concept video: rrlab.tech/projects/follow-cart/. Status: concept stage (not a hardware production demo).

What This Proves

FDE can turn scenario, form, and reachable pricing into a showable product concept—lock the vibe with stills and motion first. From utility to lifestyle: embodied AI need not stay out of reach.

02

How We Work
Every step, verifiable

01

Benchmark First

Before a single line of code, we establish an evaluation baseline. What "done" means is defined with data you approve — written into the plan. No gut-feel acceptance at project end.

02

Milestone Delivery

Weekly progress, runnable deliverables at each milestone. Unsatisfied at any stage? What's been delivered is yours, with complete documentation.

03

Research-Grade Rigor

Every architectural decision comes with stated rationale and boundaries. No "fully proven" — we say "ablation 96:76, plus consistent cross-project engineering experience." You know the evidence strength behind every conclusion.

03

Services & Pricing

01

Agent Deployment Sprint

Embedded in your team. Find the highest-value AI application scenario. From PoC to production-grade Agent system.

Deliverables: task planning & tool-calling design / Agent runtime framework / evaluation system & test suite / production monitoring plan & documentation

¥50,000 – 150,000 / project · 4–12 weeks
02

Multi-Model Orchestration Architecture

Evaluate your model strategy. Design routing, orchestration, and fallback architecture. The right model for the right task — optimal cost-capability allocation.

Deliverables: model routing & scheduling strategy / collaborative inference pipeline / cost optimization plan / performance benchmark report

¥30,000 – 80,000 / project · 3–8 weeks
03

AI Capability Benchmarking

Custom evaluation framework for your business scenario. Quantify real capability boundaries of multiple models. Deliver selection recommendations.

Deliverables: scenario-specific evaluation framework / test suite construction & execution / multi-model cross-comparison report / model selection recommendations

¥20,000 – 50,000 / project · 2–5 weeks
04

Contact
Typically responds within 24 hours

01

Email with your requirements: project background, target outcomes, timeline

02

30-minute initial alignment: confirm technical path and project scope

03

Proposal & pricing: explicit technical plan, acceptance baseline, fee structure

04

Contract & kickoff: milestone-based delivery, each stage verifiable

Email rodneyrui@gmail.com
GitHub github.com/rrlab-tech
ORCID 0009-0004-3898-4655

Use "RRLab Consulting" in your email subject for faster routing.