Open to AI Deployment / FDE roles

Shengwei You游盛巍 · Jason

AI Deployment Engineer / FDE

I build the full reliability layer around frontier model calls: multi-provider orchestration, retrieval grounding, on-chain execution, and provable safety gating. I verify every claim with seeds, logs, and receipts, and I treat my own results with a reviewer's skepticism.

Shengwei You (游盛巍)

Focus

Agent safety & deployment

Open to

FDE & AI roles

By the numbers

Verifiable, not claimed

Every figure traces to a seed, a log, and a result file. Click a number to jump to the system that produced it.

How I engineer

Production AI you can verify, not proofs-of-concept

Most AI projects die between the demo and the deploy. I build for the part after the demo: systems that do real work, fail safe under pressure, and prove it.

01

Reliability layer around frontier models

Multi-provider orchestration (OpenAI, Anthropic, local) with automatic failover, content-hashed caching, and per-call cost, latency, and token telemetry.

02

Formal verification

Safety state machines specified and model-checked in TLA+, so the kill-circuit is proven safe, not hoped safe.

03

Adversarial red-teaming

A Q-learning adversary I use to attack my own systems, with the failure modes documented honestly instead of hidden.

04

Reproducibility

Every reported number traces to a seed, an LLM trace, and a result file. One command reruns everything. I document the limits of my own evidence.

Also available for AI deployment delivery work through Lion Protocol: one scoped workflow taken from problem to a live, monitored, owned system, typically in about 90 days.

Selected work

Four shipped systems

Built solo, end to end. Each one ships with the measurements and the means to reproduce them.

OC1: Agent Safety Control System

A multi-provider agent stack with a formally verified safety circuit.

  • Built: multi-provider LLM orchestration with auto-failover, a TLA+ safety state machine, prompt-injection detection, an EVM execution layer, and a RAG policy oracle. Solo: 18,400+ lines of Python, 7 Solidity contracts, 14 test suites.
  • Measured: P95 about 3.9s with sub-5ms safety gating; 5,866,037 TLA+ states, 0 violations; prompt-injection F1 0.765 at a 5% false-positive rate; survived 100/100 adaptive RL attack episodes.
  • Verifiable: one-command pipeline, seed-controlled, with claim-to-artifact traceability.

Code available on request.

OC2: On-Chain Multi-Agent Debate

Three LLM roles argue, then the verdict is enforced on-chain.

  • Built: a 3-role debate (Proposer, Challenger, Judge) that vets DeFi actions before execution, committed on-chain via 4 Foundry-tested Solidity contracts with ECDSA verification and transcript hashing.
  • Measured: 90.2% verdict accuracy [87.6, 92.8], 95% expert agreement on real on-chain actions, 16.8s end to end, 8.7% cheaper gas than the Optimism fault-proof baseline (1.95M vs 2.13M gas).
  • Verifiable: 76 Foundry tests covering unit, adversarial, and gas benchmarks.

Code available on request.

MVF-Composer: Stablecoin Reserve Controller

Accepted and presented at IEEE ICBC 2026.

  • Built: 12 concurrent LLM agents across 4 archetypes over OpenAI, Anthropic, and DeepSeek, with Pydantic-validated outputs and a constrained mean-variance optimizer. About 12,500 lines of typed Python across 46 modules.
  • Measured: across 1,200 reproducible simulations, 62% peak peg deviation reduction and 3.1x faster crisis recovery versus the industry baseline.
  • Verifiable: peer-reviewed, with seed, commit hash, and timestamp captured per run.

Code available on request.

NOC: Cryptographically Verifiable Oracle

On-chain SNARK verification at a fraction of incumbent cost.

  • Built: a Solidity 0.8.23 oracle with on-chain Groth16 and BN254 verification, plus staking, slashing, and reputation-weighted consensus.
  • Measured: about 0.04 USD per update on L2 (about 21x cheaper than incumbent committee oracles), 100% Byzantine detection up to a 37.5% adversary fraction.
  • Verifiable: 76 passing tests, including a 256-run fuzz suite and gas benchmarks.

Code available on request.

Research

Multi-agent trust & safety

My PhD work sits at the intersection of autonomous agents, consensus design, and safety. The engineering and the research feed each other.

Calibrated multi-agent decision systems

How do you trust the output of an autonomous agent? My research develops a proposer, challenger, judge architecture, where independent agents argue and adjudicate before a decision is acted on, as a path to calibrated, defensible automation. The same chassis I study academically is the safety layer I deploy in real systems: outputs are challenged and judged, not blindly executed.

Themes: multi-agent systems, consensus and trust, calibrated forecasting, blockchain and decentralized-system safety.

Selected publications

Peer-reviewed work

2026

Hybrid Stabilization Protocol for Cross-Chain Digital Assets Using Adaptor Signatures and AI-Driven Arbitrage

Book chapter · You, Kuehlkamp, Nabrzyski · DOI: 10.1007/978-3-032-00495-6_8

2024

Persona-Preserving Reputation Protocol (P2RP) for Enhanced Security, Privacy, and Trust in Blockchain Oracles

Cluster Computing (journal) · You, Radivojevic, Nabrzyski, Brenner · DOI: 10.1007/s10586-023-04222-4

2023

Mining User Behavior in Decentralized Applications for Blockchain Trust and Security Analytics

IEEE BCCA 2023 · You, Joshi, Kuehlkamp, Nabrzyski · DOI: 10.1109/bcca58897.2023.10338860

2022

Trust in the Context of Blockchain Applications

IEEE BCCA 2022 · You, Radivojevic, Nabrzyski, Brenner · DOI: 10.1109/bcca55292.2022.9922068

Experience

Track record

From open-source infrastructure at The Linux Foundation to shipping AI agent systems today.

2024 to Present

AI Deployment Engineer (Independent)

Lion Protocol · Hong Kong

Design and ship AI agent systems: multi-provider orchestration, retrieval grounding, and safety gating, served through FastAPI with observability and human-in-the-loop checkpoints built in. Scope, build, harden, and hand over systems clients own outright.

2025 to Present

Guest Lecturer, Financial Computing

School of Business, Hong Kong Baptist University

Guest lecturer on a graduate-level financial-computing course, teaching students to build and backtest quantitative models in Python (NumPy, pandas, scikit-learn).

2019

Blockchain Full-Stack Engineer (Internship)

The Linux Foundation · San Francisco, CA

Built a real-time performance dashboard for Hyperledger Fabric using the MERN stack and authored Caliper-CLI documentation that cut developer onboarding time. Work presented at Hyperledger Global Forum 2020.

2018 to 2019

Graduate Student Researcher

Purdue University, CSE · West Lafayette, IN

Built a secure Hierarchical-Deterministic (HD) wallet in Go and open-sourced it. Implemented an adversarial ML attack on CNNs and presented findings to the university security community.

Education

Academic background

PhD Candidate, Computer Science

University of Notre Dame

Advisor: Jarek Nabrzyski

Dissertation on trust and systemic risk in decentralized systems, and multi-agent architectures for calibrated, safety-aware automation.

M.S., Computer Science (Information Security)

Purdue University

2018 to 2019

Information security, cryptography, and adversarial machine learning.

B.A., Mathematics

University of Washington

2014 to 2018

Applied mathematics with a focus on machine learning and NLP.

Tech stack

Tools I build with

Agents & AI

Multi-provider orchestrationLangGraphRAGPrompt-injection detectionEval-harness design

Engineering

PythonFastAPIasyncioPydantic v2pytest

Rigor

TLA+ model checkingBCa bootstrap CIsPre-registered experimentsReproducible pipelines

Blockchain

SolidityFoundryHardhatGas profilingEVM
Currently open to

Where I can plug in

AI Deployment Engineer / FDE

Primary focus. OpenAI, Anthropic, Crypto.com.

AI × Crypto, Innovation Engineer

Binance, OKX, Crypto.com.

DevRel / DevX (Mandarin)

OpenAI DevX, MiniMax, Mistral, Qwen, DeepSeek, Zhipu.

Also building a Mandarin-first course, "Your First AI Employee", and a weekly Skill Studio at Lion Protocol OPC. The cadence is make, document, share, sell.

Contact

Let’s build something that ships

Open to AI Deployment / FDE roles, AI and crypto engineering, DevRel, and selective consulting. I read every message.