SilkRoad Agents - Devpost Submission

Inspiration

Western investors face a $5,000+ problem: evaluating Chinese tech companies requires expensive consultants, weeks of research, and bilingual expertise that's hard to find.

When a VC firm wants to assess Unitree Robotics (宇树科技) or CATL's European expansion, they need someone who can read MIIT (工信部) policy documents, understand NDRC (发改委) regulations, and compare with Western equivalents like Tesla or Panasonic.

We asked: What if ERNIE 4.5's native bilingual capabilities could replace this entire workflow?

What it does

SilkRoad Agents is a 3-agent AI pipeline that generates professional Investment Memos in 2 minutes instead of 2 weeks:

1. Chinese Tech Scout 🇨🇳 - Analyzes domestic players, market size (in RMB), and regulatory context from MIIT/NDRC
2. Western Market Analyst 🇺🇸 - Compares with US/EU equivalents, identifies risks, rates insights
3. Investment Synthesis Manager - Generates structured Investment Memo with BUY/HOLD/PASS recommendation

Key Features:

• Upload Chinese PDFs for OCR extraction (PaddleOCR)
• Voice input for research topics
• Adjustable risk weight sliders
• Download reports as Markdown or HTML
• Real-time agent debate visualization

How we built it

The architecture leverages ERNIE 4.5's Mixture-of-Experts (MoE) design with 300B parameters (47B active) for efficient bilingual inference. We integrated it into CAMEL-AI's BaseModelBackend to enable structured agent debates.

Challenges we ran into

1. Async Event Loop Starvation - Synchronous requests calls inside async generators caused the system to hang. We migrated to httpx.AsyncClient for true non-blocking I/O.

2. PaddleOCR Initialization - The latest PaddleOCR deprecated several arguments (use_angle_cls, use_gpu, show_log). We simplified to PaddleOCR(lang="ch") after debugging multiple initialization failures.

3. Demo Data Matching - "Robotaxi" prompts were incorrectly triggering "Humanoid Robots" responses. We fixed the if/elif ordering to prioritize specific keywords like "apollo" and "driving".

4. ChatCompletion Normalization - CAMEL-AI expects OpenAI-style ChatCompletion objects, but both Novita and Baidu APIs return raw dicts. We implemented _format_response() to normalize all responses.

Accomplishments that we're proud of

• 99%+ cost reduction: $0.15 vs $5,000+ for equivalent analysis
• 40x faster: 2 minutes vs 2-4 weeks
• True bilingual output: Agents naturally include Chinese names (中文) and regulatory body abbreviations
• Production-ready: Docker deployment, health checks, error handling
• Full OCR pipeline: Upload Chinese PDFs, extract text, feed to agents

What we learned

1. ERNIE 4.5 outperforms Western LLMs for China analysis - It knows MIIT subdivision responsibilities, understands "Guochao" (国潮) branding, and naturally generates bilingual content
2. Multi-agent debate improves accuracy - Having agents challenge each other produces more balanced, nuanced memos
3. Async-first matters - SSE streaming requires careful event loop management; blocking calls kill the UX
4. PaddlePaddle ecosystem is powerful - PaddleOCR + ERNIE create a complete Chinese document intelligence stack

What's next for SilkRoad Agents

• Multi-document analysis: Compare multiple company filings side-by-side
• Real-time market data: Integrate live stock prices, news feeds
• Custom agent roles: Let users define specialized analysts (semiconductor expert, EV analyst)
• Export to PDF: Professional formatting with charts and tables
• Enterprise API: RESTful endpoints for integration into existing VC workflows
• Fine-tuned models: Train on actual Investment Committee memos for domain-specific output

Built With

• ERNIE 4.5
• CAMEL-AI
• PaddleOCR
• FastAPI
• Next.js
• Python
• JavaScript
• Docker

Built With

• camel-ai
• docker
• ernei-4.5
• fastapi
• javascript
• nextjs
• paddleocr
• python