Outcomes you can verify

Upload documents into a knowledge base, ask in natural language, and get answers that cite original sources.

Highlights

• Cited answers—source markers open surrounding context
• Multi-turn context—follow-ups stay bound to your documents
• Multi-model and private—OpenAI, DeepSeek, Ollama; data stays in domain

Customer gains

• Knowledge assets—policies, contracts, and research in one searchable store
• Low-cost build—private deployment without labeling or training
• Open API—for business systems and agents

GF Securities’ OTC derivatives AI trading-instruction module (Xinchuang) is live: IM speech-like orders, 7×24 service, trader final review.

Orders move through four stages: multi-turn intent parsing for text, images, and Excel with auto clarification; quote confirmation cards before review; three hard risk checks on permission, cash, and compliance; trader final review as the sole execution path with a full audit trail.

Highlights

• Multi-turn intent—unified understanding of text, images, and Excel
• Structured order generation—confirmation cards, not free-form model text
• Deterministic hard risk—permission, cash, compliance, then trader execution

Customer gains

• 7×24—clients can inquire and order beyond trading hours
• Speed and accuracy—96% instrument recognition; less manual re-entry
• Compliant audit—full traces; explain vs decide stay separated

Key metrics: 96% instrument recognition · 7×24 service · 3 hard risk checks · live

End-to-end client order journey

First, the client sends speech-like or multimodal orders in WeCom (text, image, Excel). Second, the system parses intent and validates business fields—instrument, side, size, price—completing missing parameters across turns and emitting a structured instruction with permission checks. Third, risk pre-check validates cash, suitability, and quotes. Fourth, a confirmation card is generated; after client confirm it enters pending review. Fifth, the trader approves and investment management or the counter executes. Sixth, fills return to the WeCom group with a full audit trail.

The LLM layer understands and plans, emitting checkable instructions. The deterministic layer owns risk, split orders, and execution—the only path to the counter. Results feed back into the data system as replayable traces.

Live since June 2022 across three service phases, covering Chancheng, Gaoming, Sanshui, Nanhai, and Shunde—driven by supervision agents.

Highlights

• Auto reports—monthly, semi-annual, and annual exports
• Smart violation detection—parking, helmets, over-fleet cleanup
• Dispatch optimization—congestion/gap and transfer analysis

Customer gains

• Supervision efficiency—from patrols to data and agents
• City-wide coverage—five districts, 125,000+ vehicles online
• Human-in-the-loop traces—AI advice, human confirm, full lineage

Key metrics: 125,000+ registered vehicles · 92.6% parking compliance · 95.4% helmet rate · 97%+ online

Agentic RAG, ~1 minute for hundred-page contracts, private deployment, full audit trail.

Highlights

• Agentic RAG—four-level retrieval with statutory cites on risk clauses
• Multi-agent orchestration—parse, retrieve, reason, reflect; human override
• Dynamic knowledge hub—statute tracking and experience without restarts

Customer gains

• 10–15× efficiency—hundred-page contracts in ~1 minute, 98%+ accuracy
• ~90% lower compliance risk—fewer misses; conclusions bound to law
• Fully private—data stays on the intranet with end-to-end audit

Key metrics: ~1 min per 100 pages · 10–15× review speed · −90% compliance risk · private deploy

Dual-track positioning: rules quantification crossed with LLM semantics—from data to allocation advice end to end.

Inputs include macros (FRED, BLS, statistical bureaus, Wind), news and research (event-driven, sentiment), and policy events (central bank, fiscal, geopolitics). Track A is rules-based Merrill 4-phase plus Pring 6-phase computation. Track B is LLM semantic judgment for events, cycle narrative, and sentiment. After cross-check, the system outputs cycle positioning (phase and evidence chain) and allocation advice (add/reduce with confidence), tied to desk holdings.

Highlights

• Rules + LLM dual track—Merrill/Pring frameworks crossed with semantics
• Fast and reproducible—macro response from hours to minutes, traced
• One-click replay—returns, win rate, and drawdown by phase for stocks, bonds, commodities

Customer gains

• +60% research efficiency—automated collection and baseline analysis
• Allocation–holdings loop—add/reduce advice against desk positions
• Data stays offline—intranet deploy on a Xinchuang stack

Key metrics: +60% research efficiency · minute-level event response · Merrill 4 + Pring 6 · dual-track check