Founding Engineer · Citronetic · Shenzhen

Gan Liu刘淦

I build the agent-harness layer between foundation models and products.

Currently building the harness behind Ockine  ·  CS @ SUSTech, graduating Jun 2026  ·  open to agent-infrastructure & applied-AI roles

Focus Agent infra · MCP · orchestration 7 production systems built Research CVPR 2026 accepted · ICRA under review Ships on Cloudflare · Next.js · Claude Code
See the work Get in touch

00 · Profile

Highlights

  1. Founding engineer at an ultra-early-stage AI startup — I've built five agent-harness / agent-infrastructure systems 0→1, on my own, end to end (seven in all, including the to-VC product Ockine).
  2. I turn messy agent capability into one stable, versioned interface (MCP + REST + a realtime session protocol), so web, native iOS, and multiple AI clients (Claude Code, Codex, Cursor) all connect to the same thing.
  3. First author on CVPR 2026 (accepted), with a second paper under review at ICRA; I follow video and image models closely.
  4. SUSTech Top-10 Outstanding Graduate.
7systems shipped
CVPR 2026paper accepted
Top 10SUSTech graduate
3.79/4.0GPA · top ~10%

01 · Selected work

Systems I've built

Four production systems in the agent-infrastructure layer — from an autonomous multi-agent dev harness to a multimodal product serving two clients from one API. Each grounded in a live repo or shipped product.

dl

dev-loop

Creator · OSS

Multi-agent harness · unified MCP contract · shared state machine · model routing

An autonomous product team that lives in your repo — 10 role-specialized agents (PM, QA, Dev, and more) that pick up tickets, implement, test, and ship, coordinating entirely through board state.

  • A single-source 23-tool MCP contract (Zod schemas over tickets, an append-only event log, versioned docs) so a direct-DB server and a daemon shim expose one identical API to every client.
  • CLI-portable via an env-based identity contract: the same agents run on Claude Code and Codex against one node:sqlite system-of-record — zero agent-body edits, fail-closed identity gate.
  • Per-agent model routing (opus for design/verification, sonnet for mechanical roles) and a graceful degradation ladder — junior → senior → human — with WAL-atomic ticket claims and orphan-reclaim crash recovery.

node:sqlite (WAL) · MCP (stdio) · Zod · Claude Code + Codex · TypeScript

Open source GitHub ↗
Aq

Aquarium

Creator

Typed WebSocket session protocol · LiteLLM routing · reconnect · guardrails

A multi-tenant control plane for deploying and managing fleets of agent instances — one typed protocol for every client, model routing across 20+ providers, and shared-nothing per-tenant isolation.

  • A unified session protocol: a typed WebSocket event model (WsMessage envelope, room-based subscriptions, instance state-transition events) and a uniform ApiResponse<T> REST envelope across 30+ route modules, all from one TypeScript type package.
  • A routing + guardrails layer over LiteLLM across 20+ providers — OAuth-provider remapping/fallback, prompt-injection detection, DLP, output filtering, four security profiles — with per-instance budget enforcement.
  • Gateway RPC resilience — versioned protocol, bounded reconnect retries, ping/pong heartbeat, reconnect-and-state-resync barriers — on a control-plane / data-plane split (Docker / Kubernetes).

TypeScript · WebSocket · LiteLLM · Docker / Kubernetes · shared type package

Live aquaclaw.ai ↗
Oc

Ockine

Founding engineer

Server-authoritative harness · 2 clients / 1 API · tiered routing + fallback · credit ledger

A multimodal AI cutscene product — upload one photo, get a cinematic clip — served to a Next.js web/PWA and a native iOS app from a single shared API. Early-stage, to-VC.

  • One shared REST API and a single /api/catalog contract serve both clients, keeping templates, quality tiers, and feature flags in sync without an App Store release.
  • Model routing over Volcengine Ark Seedance and OpenRouter with one-line Draft/Standard/Pro tier swaps and automatic graceful degradation (silent fallback on error; no-key → sandbox), so generation never hard-fails.
  • A concurrency-safe server credit ledger (per-account concurrency + daily caps, non-negative-balance constraint, refund-on-failure) as the single authority for inference cost — plus a companion-agent module with rolling-summary memory.

Next.js (OpenNext) · Expo iOS · Cloudflare Workers / D1 / R2 · Clerk · async image-to-video

In active fundraising Company ↗
AgentGEO logo

AgentGEO

Personal project

One API + MCP endpoint · multi-engine answer capture · citation extraction · open-sourced skills

A GEO (generative-engine optimization) answer layer for agents: one endpoint that queries multiple answer engines and returns real answers with citations, so agents and brands can measure how they surface in AI answers.

  • One API + MCP endpoint that queries ChatGPT, Perplexity, and Gemini and returns real answers with extracted citations — the measurement layer for AI-answer visibility (GEO).
  • Client integration shipped as open-sourced agent-side skills over the same MCP contract, so any agent plugs in with no bespoke wiring.

MCP · multi-engine capture · citation extraction · open-source skills

WIP · skills open-sourced Skills ↗

02 · Also built

More from the workbench

Five more agent-native tools and platforms; a selection.

Citron Script

A novel-to-short-drama writers' room: import a novel, and your local Claude (over MCP) turns it into a structured event graph and episode scripts. No embedded model — the app exposes tools, your agent is the brain.

MCP · local-first

MemOS

An AI memory OS for LLM & agent systems — persistent, structured memory across sessions and tools.

Agents · infra

swipi-engine

An agentic framework for end-to-end web-game creation from a single prompt — a Claude-powered generation pipeline.

GenAI · games

cli-anything-godot

A Godot 4.x CLI harness for AI agents — 15 command groups, 45+ subcommands for headless game workflows.

Tooling · games

spine-pipeline

Generate complete Spine 4.2 animations from one character image — no Spine editor in the loop.

GenAI · animation

03 · Research

Published research

First author at CVPR and co-first author at ICRA — two of the most competitive venues in computer vision and robotics. The CVPR 2026 paper is accepted; the ICRA paper is under review.

RetinalDepth surgical depth-estimation results
★CVPR 2026 Accepted Denver · top-tier CV venue

Real2Sim2Real: RetinalDepth for Depth Estimation in Posterior-Segment Ophthalmic Surgery

Gan Liu*, Bingwen Dong*, Xiaoxi Lu*, Jiang Liu

The first large-scale synthetic dataset for retinal microsurgery (44,800 stereo pairs / 896 scenes) via a Real2Sim2Real Blender pipeline — plus a Temporal Depth Variance metric — lifting monocular, stereo, and video depth foundation models on real surgical video.

First author · CVPR 2026 (accepted) · Poster ↗

SurgMASt3R-SLAM surgical scene reconstruction
★ICRA 2026 Under review Top-tier robotics venue

SurgMASt3R-SLAM: Robust Real-Time Surgical Scene 3D Reconstruction by Adapting a 3D Foundation Model

Xiaoxi Lu*, Gan Liu*, Bingwen Dong*, G. Chen, M. Gong, Y. Hu, X. Zhang, J. Liu

An end-to-end framework coupling a 3D foundation model with SLAM — frame-similarity sampling plus LoRA fine-tuning mines geometric priors for depth, pose, and dense reconstruction on texture-sparse tissue under dynamic illumination.

Co-first author · under review · Project page ↗

Research projects

Wave-CDAnet

A physics-informed neural net (3D-UNet encoder + implicit decoder) for super-resolution of nonlinear wave dynamics, constrained by Rayleigh–Bénard PDE residuals.

SciML · PINN · PyTorch

RayVision

A biomimetic underwater ROV with a neural-network target-recognition system that detects and classifies marine organisms via bounding boxes, robust to the color shift, low contrast, and noise of underwater imagery.

Robotics · CV · underwater

04 · Experience

Where I've worked

Jan 2026 – Present

Citronetic Limited — Shenzhen / Remote

Founding Engineer

  • Designed and built an agent-harness layer across seven systems between foundation models and product clients — defining unified MCP + REST contracts so heterogeneous clients integrate against one stable interface.
  • Built capability-aware model routing with automatic fallback and concurrency-safe credit/usage ledgers as the single authority for inference cost, plus cross-client identity and session continuity on Cloudflare D1.
  • Shipped an early-stage to-VC multimodal product (Ockine) serving web/PWA + native iOS from one shared API.

2023 – Present

SUSTech — Jiang Liu Lab (3D vision & medical imaging)

Undergraduate Research Assistant

  • Adapt 3D foundation models (MASt3R-family) to surgical scenes — cross-domain depth estimation and real-time SLAM reconstruction — building a Real2Sim2Real synthetic dataset and domain-adaptation framework. The groundwork behind the CVPR & ICRA papers.

Jun – Sep 2025

McGill University — Montréal (Mitacs & CSC funded)

Research Intern — Machine Learning

  • Built end-to-end temporal ML pipelines for crop-yield prediction over multi-temporal, spatial agricultural data — ingestion, feature engineering, training, evaluation — lifting prediction accuracy ~12%.
  • Developed correlation models predicting field quality from drone-derived features and expert-panel ratings, validating learned predictors against human assessments.

Jan – Jun 2025

Wego — Singapore (MENA #1 travel app)

Business Analyst · Strategy Software Engineer

  • Built the full data-to-visualization pipeline behind a company-wide strategic-planning dashboard (BigQuery / SQL · Vue.js + Tableau), cutting manual quarterly-planning effort ~3×.
  • Drove it from MVP to the quarterly-planning tool used by leadership across multiple APAC markets; presented analytics at international conferences (Web in Travel — Dubai & Japan). Placed via NUS Overseas College.

Jun – Aug 2024

Cuboid.AI — Remote

Software / AI Development Intern

  • Fine-tuned LLMs for the interior-design domain (~15% higher context accuracy) and built a real-time interactive 3D product demo in Unity3D.
  • Led market & competitor research anchoring the product spec — awarded Silver at the 4th Shaoyin Cup National AI Product Innovation Competition.

05 · Education & recognition

Education & awards

2022 – 2026

Southern University of Science and Technology (SUSTech)

B.Sc. Computer Science · GPA 3.79/4.0 (top ~10%) · Top-10 Outstanding Graduate

  • Coursework: Computer Architecture, Machine Learning, Data Analytics.

Jan – Jun 2025

National University of Singapore (NUS)

Entrepreneurship & Business Analytics — NUS Overseas College (NOC)

  • Venture creation, business analytics, product strategy, UX.

Awards

06 · Toolkit

How I build

Agent / LLM infrastructure

MCP servers (stdio / http / sse) · tool-calling · model routing & graceful degradation · multi-agent orchestration · shared-state machines · event sourcing & replay · guardrails (prompt-injection, DLP, output filtering) · memory & retrieval (FTS5 + vector, RRF, MMR) · async job pipelines

Backend / platform

Cloudflare Workers / D1 / R2 / KV · Hono · Express · Fastify · REST + WebSocket API/SDK design · versioned interfaces & backward-compatible migrations · SQLite (node:sqlite / better-sqlite3, WAL) · PostgreSQL · Clerk (JWT/JWKS) · multi-tenant isolation · Docker / Kubernetes

Languages & frontend

TypeScript / Node.js · Python · C++ · SQL · Next.js · React · Vue.js · React Native / Expo · Tailwind / shadcn · SSE

ML & research

PyTorch · 3D vision & reconstruction (SLAM, NeRF, Gaussian Splatting) · foundation-model adaptation · LoRA fine-tuning · domain adaptation · temporal ML