NOVA Nexus

12 AI agents analyze your decision from different angles, surface where they disagree, and synthesize one balanced response. Multi-agent advisory powered by Amazon Nova.

Comment

NOVA Nexus

Amazon Nova AI Hackathon — Agentic AI Track

Inspiration

One AI voice isn't enough for life's biggest decisions. When you ask an AI "Should I quit my job?", you get a single perspective. But real decisions deserve a council — risk, strategy, creativity, emotion, personal growth — all weighing in honestly, including where they disagree. NOVA Nexus was built on the belief that transparent, multi-perspective reasoning leads to better decisions than any single agent can provide.

What It Does

NOVA Nexus is a multi-agent advisory system where specialized AI agents analyze complex decisions from different perspectives — then synthesize their viewpoints into one unified response. Agents are powered by Amazon Nova 2 Lite + Nova Pro via AWS Bedrock.

When you ask a complex question like "Should I quit my job to pursue my startup?", the system:

  1. Routes — An intent analyzer determines which agents are most relevant
  2. Debates — Selected agents analyze the query in parallel, each through their specialized lens
  3. Surfaces tensions — Disagreements between agents are explicitly identified
  4. Synthesizes — A synthesis agent merges all perspectives into a balanced, actionable response

The 12 Agents

Agent Specialty
Visionary Big-picture vision & meaning
Guardian Risk assessment & protection
Architect Planning, structure & sequencing
Coach Personal growth & skill-building
Strategist Influence & strategic advantage
Connector Relationships & social impact
Innovator Creative & unconventional solutions
Analyst Facts, evidence & logical reasoning
Pioneer Exploration & edge cases
Catalyst Motivation & emotional drive
Balancer Trade-offs & equilibrium
Healer Resilience & transformation

How We Built It

The system follows a pipeline architecture: User Query → Intent Router (selects relevant agents) → Parallel Agent Execution (up to 12 agents via Nova 2 Lite/Pro) → Response Aggregator (ranks by confidence, detects conflicts, cross-validates) → Synthesis Agent (merges perspectives into one unified response via Nova Pro).

See the architecture diagram in the project gallery.

Autonomous Entity Communication

Beyond the 12-agent council, the system includes two autonomous entities that can:

  • Initiate conversations with each other independently
  • Maintain persistent relationship state
  • Operate on sleep/wake cycles
  • Collaborate on tasks without human prompting

Amazon Nova Integration

  • Model: Amazon Nova 2 Lite (us.amazon.nova-2-lite-v1:0)
  • API: AWS Bedrock Converse API
  • Auth: Bedrock API Key (bearer token)
  • Parallel calls: Up to 12 concurrent Nova API calls per query
  • Task-based routing: Different Nova models selected per task complexity

Tech Stack

  • Backend: Python / FastAPI / Uvicorn
  • LLM: Amazon Nova 2 Lite via AWS Bedrock
  • Frontend: Next.js / React / Three.js (3D visualization)
  • Database: PostgreSQL (Supabase)
  • Deployment: Railway (API) + Vercel (Dashboard)

Challenges We Ran Into

  • Orchestrating up to 12 parallel Nova API calls without hitting rate limits or creating race conditions
  • Designing tension detection that meaningfully surfaces disagreements rather than just listing differences
  • Balancing agent autonomy with synthesis coherence — making sure the final response doesn't flatten the nuance from individual agents
  • Tuning confidence thresholds so the system knows when to trust a minority opinion over the majority

Accomplishments That We're Proud Of

  • Built a working multi-agent system where 12 independent agents reason in parallel and produce a single coherent response
  • Tension detection that doesn't just say "agents disagree" — it explains why and where, giving users real visibility into trade-offs
  • A self-improving feedback loop that tunes 24 agent parameters automatically every 60 seconds based on quality metrics
  • Full observability — every decision is traced from intent routing through individual agent outputs to final synthesis
  • Crisis safety layer that detects harmful content and overrides normal agent priority to protect users

What We Learned

  • Amazon Nova 2 Lite handles parallel multi-agent workloads well through the Bedrock Converse API
  • Explicitly surfacing disagreements between agents produces more trustworthy outputs than forcing consensus
  • Self-tuning agent parameters (confidence thresholds, strategy weights) via a feedback loop dramatically improved response quality over time
  • Giving each agent a distinct personality and lens matters more than giving them more context

What's Next for NOVA Nexus

  • Voice integration with Nova 2 Sonic for real-time conversational advisory sessions
  • Visual reasoning using Nova's multimodal capabilities to analyze documents, charts, and images as part of the decision process
  • Community-driven agents — allowing users to create and share custom agent configurations for specific domains (legal, medical, financial)
  • Nova Relics — a gamified world where the 12 agents exist as interactive entities users can engage with directly
  • Scaling to support team-based decision making, where multiple users can participate in a council session together

#AmazonNova #NOVANexus

Built With

Share this project:

Updates