SIBYL SYSTEM – Ministry of Welfare Public Safety Bureau

🚔🚨"A Qwen-Cloud AI-powered governance and public safety agent system that monitors, analyzes, and enforces welfare policies with scalable multi-agent collaboration." 👮🏻👮🏻‍♂️🚓

Comment

Inspiration

SIBYL SYSTEM was born from a simple, urgent question: how can AI help governments and social services act faster, more fairly, and more transparently when people’s welfare and safety are at stake?
We were inspired by real-world gaps in cross-agency coordination, slow manual triage of welfare cases, and the potential for multi‑agent systems to combine domain expertise, persistent memory, and human oversight into a single operational workflow. The project blends civic responsibility with engineering curiosity: building an agent that augments—not replaces—human decision makers.

What it does

SIBYL SYSTEM is a modular multi‑agent platform for welfare and public safety workflows. Key capabilities:

  • Continuous monitoring of incoming reports, alerts, and case updates.
  • Autonomous triage that classifies urgency, suggests next steps, and routes cases to the right team.
  • Persistent memory that stores user preferences, case history, and policy exceptions to improve decisions across sessions.
  • Human‑in‑the‑loop checkpoints for high‑risk or ambiguous cases.
  • Audit trails and explainability so every recommendation includes the rationale and data sources.

How we built it

Architecture overview

  • Frontend: responsive web UI for case review and dashboards.
  • Backend: microservices that host agent orchestration, memory store, and policy modules.
  • Memory layer: a hybrid store combining vector embeddings for semantic recall and a relational DB for structured case metadata.
  • Agents: specialized agents (ingest, triage, policy, outreach) that communicate via a message bus and negotiate task ownership.
  • Cloud deployment: backend services run on cloud instances with logging, monitoring, and CI/CD pipelines.

Technical highlights

  • Memory retrieval: we use approximate nearest neighbor search for semantic recall to keep retrieval fast; naive retrieval is (O(n)) while indexed retrieval is approximately (O(\log n)).
    [ \text{Naive retrieval: } O(n) \quad\text{Indexed retrieval: } O(\log n) ]
  • Safety and audit: every agent action is logged with a human‑readable justification and a cryptographic hash for tamper evidence.
  • Integration: REST and webhook endpoints let SIBYL connect to existing case management systems and notification channels.

Submission notes
We prepared the repository, architecture diagram, and a short demo video to meet the hackathon submission requirements. devpost.com

Challenges we ran into

  • Memory design tradeoffs: balancing recall accuracy, storage cost, and privacy constraints required iterating between dense embeddings and structured metadata.
  • Multi‑agent coordination: designing robust negotiation and conflict resolution so agents don’t duplicate work or contradict each other took several protocol revisions.
  • Explainability under token limits: producing concise, transparent rationales for recommendations while staying within model token budgets forced us to compress reasoning into structured templates.
  • Deployment constraints: ensuring graceful degradation when cloud connectivity is poor (EdgeAgent scenarios) required building local fallback logic and lightweight models for critical checks.

Accomplishments that we're proud of

  • Working multi‑agent pipeline: agents reliably decompose tasks, assign roles, and complete end‑to‑end workflows in our demo scenarios.
  • Persistent memory with privacy controls: we implemented selective forgetting and consented memory scopes so sensitive data can be redacted or expired automatically.
  • Human‑in‑the‑loop UX: reviewers can accept, modify, or reject agent recommendations with one click; every change updates the memory and audit log.
  • Polished demo and documentation: complete repo, architecture diagram, and a 3‑minute video that clearly show the system in action.

What we learned

  • Designing for humans first: automation is only useful when humans trust it; small UX details (clear rationales, easy overrides) dramatically increase acceptance.
  • Memory is a product problem, not just an engineering one: deciding what to remember, for how long, and how to surface it requires policy thinking as much as technical work.
  • Multi‑agent systems need explicit conflict resolution: without clear arbitration rules, agents can thrash or produce inconsistent outputs.
  • Operational constraints matter: latency, token budgets, and deployment targets (cloud vs edge) shape architecture choices early and often.

What’s next for SIBYL SYSTEM – Ministry of Welfare Public Safety Bureau

  • Policy tuning and pilot deployments: partner with a local welfare agency for a controlled pilot to validate real‑world impact and gather human feedback.
  • Advanced memory policies: add adaptive forgetting, consented sharing across agencies, and differential privacy options.
  • Agent marketplace: enable third‑party modules (e.g., legal advisor, mental‑health triage) to plug into the agent society with clear capability contracts.
  • Edge resilience: optimize lightweight local agents for offline operation and secure sync when connectivity returns.
  • Open source and community: publish the core repo with clear docs and reproducible demos so other builders can extend SIBYL for different civic domains.

Built With

Share this project:

Updates