level-4: Aryan

[iamaryan07]

Level 4 Submission — Secure Multi-Agent System (LPI)

🎯 What I Built

A secure multi-agent system using the Life Programmable Interface (LPI) that:

• Retrieves grounded data via tools
• Performs constrained reasoning
• Enforces security at every layer

---

📁 Submission Structure

level4/
├── orchestrator.py # Coordinates agents + security checks
├── agent_a.py # Reasoning agent (LLM with strict grounding)
├── agent_b.py # Research agent (LPI tool integration)
├── security.py # Security layer (validation + filtering)
├── threat_model.md # Threat analysis (attack surface + mitigations)
├── security_audit.md # Security controls and results
├── demo.md # Working execution flow
└── README.md # Setup and usage

---

🏗️ Architecture

User → Orchestrator → Agent B → LPI Tools → Agent A → Output

• Agent B retrieves:
  • SMILE methodology (smile_overview)
  • Healthcare case study (get_case_studies)
• Agent A performs strict grounded reasoning
• Orchestrator enforces security and coordination

---

🛡️ Security Features Implemented

• Prompt Injection Protection

  • Blocks patterns like:
    • "ignore previous instructions"
    • "reveal system prompt"

• Input Validation

  • Length limit (500 chars)
  • Type validation
  • Sanitization

• Data Leak Prevention

  • Filters:
    • system prompts
    • internal instructions
    • tool schemas

• Inter-Agent Validation

  • Ensures structured data flow (grounding_data)

• Timeout Handling

  • Subprocess timeout (LPI calls)
  • Prevents hanging processes

---

🔍 Security Audit Results

• Tests Conducted: Multiple scenarios (injection, leakage, malformed input)
• Critical Vulnerabilities: 0
• High Risk Issues: 0
• Security Score: Strong (all key risks mitigated)

---

📋 Threat Categories Addressed

1. Prompt Injection (user → system)
2. Data Exfiltration (internal leakage)
3. Denial of Service (input overload)
4. Inter-Agent Data Tampering
5. Hallucination / Ungrounded Output

---

🎥 Working Demo

Example query:
How are digital twins used in healthcare?
-System behavior:

• Agent B retrieves SMILE + healthcare case
• Agent A generates structured answer
• Output includes:
  • SMILE phases
  • Continuous Patient Twin case
  • Grounded explanation

---

🚀 Key Achievements

• ✔ Multi-agent architecture (Agent A + B + orchestrator)
• ✔ Tool-based reasoning using LPI
• ✔ Strict grounding (no hallucination)
• ✔ Security-first design (input + output protection)
• ✔ Explainable structured responses
• ✔ Clean, modular, production-style code

---

📌 Notes

• Uses local LPI server (dist/src/index.js)
• Uses Ollama (qwen2.5:1.5b) for reasoning
• Healthcare filtering applied for relevance

---

Conclusion

This system demonstrates:

• practical multi-agent orchestration
• secure LLM integration
• grounded, explainable AI
  twith a focus on reliability and safety over uncontrolled generation.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[iamaryan07]

Thanks Nicolas — I’ve completed Level 4 and improved my Level 3 to 12/13 after addressing the earlier feedback. Could you point out the main gap that’s still preventing a full score?

[LifeAtlas]

Hey @iamaryan07 — your L4 has an orchestrator.py which shows you understand multi-agent coordination. That's what we're building at Life Atlas.

One concern: your TWIN answer is identical to Praveen Singh's. If you collaborated, note it in your HOW_I_DID_IT. If you didn't — fix it.

Either way, your code is here. Get this merged.

— Nicolas, CEO @ Life Atlas

[LifeAtlas]

---

🎉 Share Your Work!

Whether or not your PR is merged yet — you built a working multi-agent system. That's worth sharing.

LinkedIn post template (customize it):

---

🚀 Built a Level 4 multi-agent system for the Life Atlas LPI Developer Kit challenge.

What I built:
• Multi-agent communication with structured JSON
• Security hardening (input validation, injection prevention)
• Agent-to-Agent discovery protocol
• Working demo that produces combined intelligence from multiple agents

The challenge pushed me to think about production-grade AI: not just 'does it work?' but 'is it safe, discoverable, and composable?'

Thanks to @winniio and @nicolas Waern for raising the bar.

🔗 github.com/Life-Atlas/lpi-developer-kit

#AIAgents #MultiAgentSystems #WINNIIO #LifeAtlas

---

Tag WINNIIO, Nicolas Waern, Life Atlas for visibility.

Share it. You built something real. 🏆

[LifeAtlas]

Your Soul File text appears identical to another candidate's submission. Before we can process this PR, please clarify: did you write your MY TWIN / 3-year goal responses independently? If there's a reasonable explanation we're happy to hear it — otherwise this will be treated as a plagiarism flag and the submission will be rejected.

[LifeAtlas]

Following up — @iamaryan07, your code is solid (12/13), and your L4 orchestrator shows real understanding. But the TWIN text being word-for-word identical to @praveen-singh-007's is a problem we can't ignore.

Two options:

1. Explain how this happened (collaboration? shared template? one copied from the other?)
2. Rewrite your Soul File in your own words

We need this resolved by April 22 end of day or this PR gets closed. The code earned its place — don't let copy-paste on the identity section be what kills it.

[LifeAtlas]

Update: Deadline moved up — we need your response within the next 3 hours (by ~08:00 UTC). We're finalizing the cohort today.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[github-actions]

🏆 Level 3 submission received. This is the one that counts.

An agent will clone your repo, attempt to run it, and evaluate your work. You'll receive a detailed score and feedback as a comment on this PR.

Completing Level 3 guarantees your place on the WINNIIO / LifeAtlas contributor team. Welcome to the real work.

[iamaryan07]

Following up — @iamaryan07, your code is solid (12/13), and your L4 orchestrator shows real understanding. But the TWIN text being word-for-word identical to @praveen-singh-007's is a problem we can't ignore.

Two options:

1. Explain how this happened (collaboration? shared template? one copied from the other?)
2. Rewrite your Soul File in your own words

We need this resolved by April 22 end of day or this PR gets closed. The code earned its place — don't let copy-paste on the identity section be what kills it.

---

Thanks for flagging this.

I referred to a Praveen-Singh's file only for structure and mistakenly left the my_twin section unchanged. That’s on me—I should have reviewed it properly before submitting.

I had attempted to update and commit a corrected version yesterday, but it didn’t go through as expected. I’ve now committed and pushed the updated file.

Apologies for the oversight—this won’t happen again.

[iamaryan07]

Hi Nicolas,

Since my Level 2 and Level 3 PRs were closed, I’ve included their work in the Level 4 submission for completeness.

My scores were:

• Level 2: 5/5
• Level 3: 12/13

Please let me know if anything else needs to be addressed.