Public case study
BioPulse
生命科学 Agent-native 工作台
An agent-native workbench where agents do the reading, filling and cross-auditing.
Status Public repository with MIT license detected by GitHub. README claims describe capability; no benchmark or accuracy figure is published.
What it is
BioPulse is an AI workbench for life-science research: literature, data and process in one place, with agents taking on the repetitive reading and cross-checking.
The public tree holds a cloud service, a set of independent agent packages, layered safety guardrails, vector memory with RAG, and four clients.
Why it matters
An agent that can act — fill a form, flag an anomaly, propose a plan — needs firmer constraints than one that answers a question. In a regulated setting the acceptable failure mode is silence, not confident invention.
That is why the guardrails sit beside the agent runtime rather than after it.
Current public status
Public repository, with an MIT licence detected by GitHub — the only licence on this site confirmed by a machine.
This site has not run the repository’s test suite and claims no pass rate.
What is publicly available
- The
BioPulserepository — Python, FastAPI and LangGraph, plus a React admin client, a web client, a Flutter mobile app and a WeChat mini program. - The
tests/directory, covering the agent runtime, the individual agents, services, frontend and mini program, including dedicated guardrail cases. - The licence: MIT.
Notes
No accuracy figures, customer names or revenue data are published here — none of them can be verified from outside. The README describes capabilities, not results.
Implementation status
A public FastAPI cloud service with seven agent packages, a four-surface front end and an MIT license detected by GitHub.
What exists today, and how you would check it. Nothing below is a rendered mockup of something that does not exist.
- Public code
- Public repository
| Component | State | Note |
|---|---|---|
| Agent runtime | Built | `cloud/app/agent_runtime` implements the execution loop, retry and circuit-breaker behaviour, a dead-letter queue, error taxonomy, metrics and a tool bridge. |
| Seven agent packages | Built | `agents/` holds compliance monitor, anomaly analysis, sales suggestion, sales coach analyst, knowledge worker, opportunity scanner and competitor crawler. |
| Safety guard layers | Built | Layered safety profiles and guards under `cloud/app/agents`, with dedicated L1 / L3 tests in the public suite. |
| Memory and retrieval | Built | Vector memory and RAG ingestion modules are present, with tests in `tests/agent_runtime`. |
| Compliance and inference engines | Built | `cloud/app/compliance` and `cloud/app/analysis` plus `cloud/app/eval` implement the cross-audit and inference-evaluation side described in the README. |
| Four product surfaces | Built | `frontend/` (React management console), `web/`, `mobile_app/` (Flutter) and `weapp/` (WeChat mini program) are all present in the tree, with `deploy/` for Nginx + TLS. |
| Test suite | Built | `tests/` covers the agent runtime, the agents themselves, services, the front end and the mini program, plus `cloud/tests`. This site does not re-run them, so no pass rate is claimed. |
| License | Validated | MIT, detected by GitHub and stated in the README — the only project in this portfolio whose license is machine-confirmed rather than unresolved. |
| Real tenant deployment | Not public | `tenant_configs/` and `deploy/` exist in the tree, but no running deployment or customer reference can be checked from outside. |
| Model accuracy claims | Planned | The README describes an inference engine with confidence intervals and back-testing, but publishes no benchmark. No accuracy number is claimed here. |
How to read these states 4
- Built Implemented in code. Says nothing about how good it is.
- Validated Built and checked against a specific artifact we can point at.
- Planned Identified for future work. Nothing to inspect yet.
- Not public Reported to exist, but no public artifact can be checked.
Repository index
BioPulse/
├── cloud/ FastAPI service (953 files)
│ └── app/
│ ├── agent_runtime/ execution loop, retry, circuit breaker, tool bridge
│ ├── agents/ layered safety profiles and guards
│ ├── analysis/ inference and causal chain evaluation
│ ├── compliance/ cross-audit and flow verification
│ ├── eval/ evaluation harness
│ ├── repositories/ persistence layer
│ ├── routers/ REST surface
│ ├── schema/ typed request / response models
│ └── migrations/
├── agents/ 7 packages: compliance, anomaly, sales suggestion,
│ sales coach, knowledge, opportunity, competitor
├── frontend/ React 18 + Vite management console
├── web/ React web client
├── mobile_app/ Flutter (Android / iOS, offline-first)
├── weapp/ WeChat mini program
├── tests/ agent_runtime, agents, services, frontend, weapp
├── deploy/ Nginx + TLS, backup, CI / CD
└── LICENSE MIT (detected by GitHub)One cloud service, four client surfaces, and the agent packages beside them — 1,297 Python files in total.
Source GitHub API — huangdi97/BioPulse tree listing, read 2026-09-18.
Documented entry points
python -m venv venv && source venv/bin/activate
pip install -r cloud/requirements.txt
cp .env.example .env # DEEPSEEK_API_KEY etc.
python -c "from cloud.app.database import init_db; init_db()"
uvicorn cloud.app.main:app --reload --port 8000
curl http://localhost:8000/agent-gateway/execute \
-H "Content-Type: application/json" \
-d '{"agent_key":"analyst","goal":"分析最近的拜访数据"}'
python3 -m pytest cloud/app/tests/ -vQuoted verbatim from the README so the commands on this page match the repository.
Source BioPulse README.md — quoted verbatim.
What this page does not claim
No deployment, customer, revenue or accuracy figure is published here. The README states production readiness in infrastructure terms — Prometheus, Nginx, backups, CI — which is exactly how far this page goes.
Source Site editorial standard.