Projects 项目展示

A cloud infrastructure grant from Tencent Cloud International, supporting latency-sensitive quantitative trading and agent workloads on low-latency APAC nodes interconnected over Tencent's private backbone.

What the infrastructure provides

• Low-latency APAC nodes: high-availability regions in Hong Kong and Singapore, physically close to major exchanges
• CCN private-backbone interconnect: Cloud Connect Network links regions over a private backbone rather than the public internet, keeping cross-region latency stable and free of public-BGP jitter
• Financial-grade compliance: SOC 2 Type II, ISO 27001 and PCI DSS, plus regional certifications (Hong Kong ICAR, Singapore MTCS)

How I plan to use it

• Latency and jitter benchmarking of a trading loop across APAC regions, measured against a public-internet baseline
• Backtesting and agent-rollout runs that extend the Quant Trading Agent built at AQUMON — LangChain orchestration over Futu OpenAPI real-time data and automated execution
• Serving experiments for agentic post-training workloads where decision latency, not throughput, is the binding constraint

Grant support

Program	Tencent Cloud Quant Infrastructure Grant
Support	USD 1,000 cloud credits
Provider	Tencent Cloud International
Focus	Low-latency quantitative trading and agent workloads across APAC

An open-source project on GitHub (richardChenzhihui/OfficeBuddy, 53★ · 6 forks): an agent that edits Word/Excel files from natural language and then proves each edit — the real Microsoft Office application re-renders the document, and an independent multimodal verifier audits the rendered pages before the loop moves on. Vision in the loop, not blind file writes.

Why it maps to agentic post-training

• The harness is a verifier-in-the-loop system: plan → act → render → independent structured verdict → targeted repair — the same shape as reward design and eval loops in agent RL post-training
• Failure-mode engineering: normalized error signatures, forced strategy switch after repeated same-class failures, hard per-step / per-task budgets
• Safety by construction: works on isolated copies only, byte-level snapshots with undo before any write, document content treated as data (prompt-injection defense)

Engineering highlights

• Rendered-screenshot verification: AppleScript drives Word/Excel to export PDF → page images → pixel-level diff with red-box annotations marks exactly what changed → a stateless visual verifier returns a structured verdict (no access to edit history, so it cannot rationalize failures)
• Verified-baseline ratchet: diffs are taken against the last passed render, so regressions cannot silently become the new normal
• Error-escalation ladder: retry → switch strategy → ask the user, under hard budget ceilings
• Model-agnostic core: hot-swappable backends (Kimi-K2, MiniMax-M3, OpenAI/OpenRouter-compatible endpoints)
• Benchmarked: a 9-task Render-Truth Bench (108 runs) shows substantially higher defect-awareness than an OfficeCLI baseline

Traction

• 53★ and 6 forks on GitHub — an actively used open-source project with an active issue/feedback loop from real users

A data-and-model framework for trustworthy medical deepfake detection, built around evidence-grounded reasoning, forgery localization, and a public demo stack.

What is included

• ACL 2026 main conference paper on interpretable medical forgery detection
• MedForge-90K dataset: 30K real images, 30K lesion implant forgeries, and 30K lesion removal forgeries
• MedForge-Reasoner: a Qwen3-VL based detector using a Localize-then-Analyze reasoning pipeline
• Interactive demo for medical image deepfake detection and reasoning visualization

Why it matters

• Moves beyond black-box real/fake prediction to localized, evidence-grounded explanations
• Targets realistic lesion implantation and removal risks in chest X-ray, brain MRI, and fundus images
• Combines dataset, model, and demo into a single research artifact instead of a paper-only release

Public resources

Asset	Details
Paper	ACL 2026 Main Conference
Dataset	MedForge-90K, covering CT, MRI, and X-ray with 19 lesion types (50K+ Hugging Face downloads)
Model	MedForge-Reasoner on Hugging Face
Demo	Online detector Space for interactive testing

The dataset behind the EMNLP 2026 Main Conference paper Med-Banana: Learning Agentic Quality-Controlled Medical Image Editing from Success-and-Failure Trajectories: 50,635 successful edits and 37,822 failed attempts with full logs across three modalities and 23 disease types — failure trajectories as agentic post-training supervision, with 100K+ downloads on Hugging Face.

Key Features

• 50,635 successful edits across 3 medical imaging modalities
• Chest X-ray: 12 pathology types (Pneumothorax, Pleural Effusion, etc.)
• Brain MRI: 4 tumor types (Glioma, Meningioma, Pituitary)
• Fundus photography: 7 disease types (Diabetic Retinopathy, Glaucoma, etc.)
• Bidirectional editing: lesion addition and removal
• LLM-as-Judge quality control with medically grounded rubric
• 37,822 failed attempts with full conversation logs for preference learning and alignment research

Dataset Statistics

Modality	Task	Diseases	Success	Failed
Chest X-ray	Add	12	9,854	7,971
Chest X-ray	Remove	12	10,667	4,750
Brain MRI	Add	4	4,536	8,630
Brain MRI	Remove	4	4,355	6,949
Fundus	Add	7	18,505	3,162
Fundus	Remove	7	2,718	6,360
Total		23+	50,635	37,822

A compute-supported project from MiniMax, focused on turning long-context, multimodal, and Agent capabilities into medical foundation model development and clinically verifiable workflows.

Project focus

• Medical foundation model development: long-context, multimodal, and Agent capabilities for healthcare workflows
• Data flywheel: evaluation findings are converted into targeted cases, feedback signals, and iterative improvement loops
• Clinically verifiable Agent workflows: outputs are structured around traceable evidence, review checkpoints, and reproducible decision paths

Grant support

Program	MiniMax Cowork Team Fellowship
Support	USD 4,500 compute grant
Direction	Medical foundation model development and clinically verifiable Agent workflows

A zero-shot detection framework for identifying LLM-generated text in specialized domains like medicine and law, using normalized entropy-based scoring and domain knowledge distillation.

Key Innovations

• Zero-shot detection: No training data required for new domains
• Normalized entropy scoring: Robust metric for specialized text
• Domain knowledge distillation: Leverages domain-specific patterns
• Cross-domain robustness: Tested on medical, legal, and financial texts

Performance Highlights

Metric	Improvement
AUROC	+14.4% vs. SOTA
Recall @ 0.1% FPR	+64.0% vs. SOTA
Zero-shot Capability	No training needed

Applications

• Detecting AI-generated medical content to combat misinformation
• Verifying authenticity of legal documents and contracts
• Ensuring integrity in academic and scientific publishing
• Quality control for financial reports and analysis

A GPU-accelerated legal-domain speech-to-text service delivered for Haiwen & Partners LLP (HK), built on Whisper Large-v2 and packaged as a production serving stack.

Serving stack

• Whisper Large-v2 with GPU acceleration for legal-domain audio
• Docker + FastAPI serving pipeline for reproducible deployment
• 7.8% average WER on the delivered legal transcription workload

An autonomous trading agent built at AQUMON on a LangChain architecture, orchestrating market analysis, signal generation, decision-making, and execution monitoring into a single closed loop for programmatic Hong Kong equity trading.

What it does

• LangChain orchestration connecting market understanding, signal generation, strategy decision, and execution monitoring
• Futu OpenAPI integration for real-time market-data streaming and automated order execution
• Strategy validation loop to speed up iteration and deployment of programmatic strategies

Why it matters

• Demonstrates agent orchestration and tool-calling against a real, latency-sensitive external API
• End-to-end loop from perception to action, the same shape as agentic post-training environments