DI Advisor: regulated multi-agent advice

Most investing chatbots will tell anyone anything. Under Malaysian securities law (SC Technical Note 1/2022), digital investment advice is a regulated activity, so a system that just answers prompts is a liability, not a product. DI Advisor is a multi-agent adviser where the regulated decision stays deterministic and the agents sit at the periphery. Three specialist agents collaborate through Band: - Interviewer: maps a client's free-text answers (English or Malay) onto a structured suitability questionnaire and confirms them back. - Explainer: turns the engine's result into plain language, with every percentage and ringgit figure verified against an engine fact before it is shown. - Compliance Q&A: answers follow-up questions from approved materials only, refusing and escalating when a question falls outside the grounded corpus. Between them sits a deterministic engine that scores risk, runs an unsuitable-investor gate (it can cap the risk level or decline service outright), builds a model portfolio, and attaches mandatory disclosures. No agent can reach that decision; the host invokes it. Band is the collaboration layer. Agents pass structured context, hand off work in a fixed order, and route to a licensed human on any decline, out-of-scope question, or low-confidence answer. Every agent turn, tool call, and decision is written to a hash-chained audit log you can verify end to end. This is Track 3: review, traceability, escalation, and careful decision-making are the product, not an afterthought. A refusal is a success state. Built zero-dependency; runs on demo personas only. Not investment advice; a deploying entity needs SC authorisation.

Dently - Dental AI Scribe

SkillPath AI: Multi-Agent Workforce Development

SkillPath is an AI-powered Enterprise Workforce Reskilling Orchestrator that helps organizations identify skill gaps, create personalized learning journeys, monitor employee progress, and generate HR-ready compliance reports through a coordinated multi-agent system. The platform uses five specialized AI agents working together through Band rooms. The Gap Analyst compares employee profiles with job requirements to identify missing skills. The Curriculum Architect creates customized learning plans using AI-powered recommendations. The Coach Agent provides continuous guidance and daily learning support. The Progress Tracker monitors course completion, assessment scores, and engagement metrics, while the HR Reporter generates audit-ready reports, certification summaries, and workforce readiness insights. Band serves as the collaboration layer where agents share context, coordinate tasks, escalate risks, and involve HR managers when approvals are required. Each employee can have a dedicated learning room, enabling personalized coaching while maintaining organization-wide visibility. By automating workforce development and compliance tracking, SkillPath reduces manual HR effort, improves employee growth, increases training effectiveness, and helps enterprises build a future-ready workforce through transparent, AI-driven decision making.

FreshMart: A Multi-Agent Grocery Experience

FreshMart is a full-stack grocery ordering platform built on a multi-agent architecture using the Band SDK. Three specialized AI agents work together behind the scenes: a Manager agent that routes requests, a Query agent (LangGraph) that answers questions about products and orders, and an Action agent (CrewAI) that handles order mutations like cancellations and updates. On the frontend, customers get a polished Next.js portal where they can browse 30+ products with a shopping cart, place orders, and track them through the full lifecycle pending, processing, shipped, delivered. A built-in AI chat assistant (powered by Claude) lets customers view their orders, change quantities, or cancel — all in plain conversation. Key technical highlights: - Band SDK multi-agent orchestration across three frameworks (Anthropic SDK, LangGraph, CrewAI) - Supabase for auth, database, and real-time data - Background job that automatically advances order statuses over time - Cookie-based auth with scoped data access so each customer only sees their own orders - Full cart persistence, stock management, and server-side checkout validation

Council production layer for LLM Council pattern

Andrej Karpathy posted **llm-council** on November 22, 2025. 20.8K stars. **VentureBeat called it "the missing layer of enterprise AI orchestration."** **Council is that production layer.** Five archetypes (Founder / Strategist / Security / Skeptic / User Advocate — swappable) answer the same decision **in parallel**. Each returns a structured verdict with confidence + a falsifiable prediction. A sixth agent — Claude Sonnet 4.6 as Oracle — reads all five and weighs minority opinions the majority is dismissing. The output is not "the average." It is: **"4/5 say ship, 1 says wait — here is why the wait matters."** When the user resolves the decision in 7 days, Council Brier-scores each voice so over months you learn which voice's no's were load-bearing and which were noise. **Calibration is the moat.** **Why Regulated & High-Stakes Workflows**: regulated workflows are the failure mode this product was built for. Single voice, single confidence, no audit, no calibration is the pattern that moves money in minutes. Council is the verification layer. Honest disclosure v0.4.0 does not wrap the Band SDK. Slack-native surface ships July 13 (Slack Agent Builder Challenge); Band-coordinated multi-MCP is surface 2-N on the roadmap. Production-grade for what shipped over roadmap-grade for what is claimed.

MedBand

MedBand is a human-in-the-loop healthcare workflow MVP built by The Billionaire Republic for the Band of Agents Hackathon. The problem is simple: healthcare service requests often move slowly because intake, verification, resource availability, and final approval happen across disconnected people and tools. MedBand solves this by using Band AI agents to coordinate the workflow in a visible, auditable Band room. Our working MVP proves one complete Pharmacy workflow. A patient submits a request through the web form. The Coordinator opens the case, Intake structures the patient request, Verification checks whether the request can proceed, Resource confirms availability at the selected institution, and Approval Desk prepares the case for human review. The final decision is not made by AI. A human reviewer manually approves or rejects the case. Verified proof case: MEDBAND-WEB-989E388C moved from web submission to CASE APPROVED at Peaceway Pharmacy. The request, BODY PAINS and PARACETAMOL, was preserved through the workflow. No AI approved this case. Final approval was made by the human reviewer. MedBand’s tagline is: Getting people to the right care, faster.