Agent core
Single path: Planner → Tools → Evidence → Report. Not a multi-role equity desk.
Evidence-first · Chinese bonds
BondLens
An Evidence-First Bond Analysis Agent for Chinese Market Data
BondLens turns a natural-language bond question into an auditable run: live/snapshot/static data, deterministic tools, optional LLM narration under numeric/language guardrails, and a reviewer-facing Trust Layer.
Not a multi-agent equity research desktop. A claim-level evidence agent for Chinese bonds.
数字由代码计算，叙述可由大模型辅助，每次输出都可追溯。 不是让 LLM 自由猜金融结论，而是工具先出证据，Agent 再编排、解释、裁决与诚实回退。
Example Runs (no API key)
Open in browser — no server required. Same idea as FinRobot-style example reports, scoped to Chinese bonds.
Codebase Snapshot
7 deterministic tools
search / describe / rank / outliers / peer / monitor / bond report — numbers never invented by the model.
Trust layer + evals
Guardrail · judge · Trust score · Evidence Pack · replay · pytest ~110 · agent 10/10 · red-team 3/3.
Core workflow
BondLens keeps the financial reasoning path visible and auditable.
Data resolver: AkShare live → cached snapshot → local Excel
Planner: overview / search / ranking / outlier / monitor / bond report / advisory policy
Deterministic tools produce structured evidence before any optional LLM call
Numeric + language guardrails decide whether model text can be final
Trust score, Evidence Pack, replay, agent evals and red-team evals
Why it matters for interviews
Agent engineering
Planner, tool routing, evidence ledger, deterministic fallback, OpenAI-compatible local LLM path.
Financial data boundary
Honest live/snapshot/static lineage; no fake ratings, credit events, or trade instructions.
Safety evaluation
Blocks unsupported numbers, investment-advice language, guarantees; CI agent + red-team suites.