Inspiration
Modern operational systems generate massive amounts of data across fragmented multi-channel and multi-node environments, but they rarely connect analysis directly to execution. Most AI solutions stop at generating dashboards. We were inspired to build a closed-loop architecture—a dynamic "OS"—that bridges the gap between AI reasoning and real-world execution, ensuring that decisions are not just suggested, but safely automated.
What it does
ConverterOS is a dynamic control operating system powered by Amazon Nova. It ingests complex operational data, builds contextual understanding through a Graph-RAG approach, and uses AI to make real-time decisions. It features a strict safety layer (ConflictResolver and safeMode) that verifies physical constraints before sending automated execution commands to edge devices or facility systems.
How we built it
*AI Engine: Amazon Nova via AWS Bedrock serves as the core reasoning engine. *Backend: Java 17 and Spring Boot 3.5 manage the API gateway, event normalization, and state management. *Frontend: React and Tailwind CSS power a Bento Grid UI that displays real-time analysis, AI decision logs, and metric-centric KPI cards. *Architecture: We implemented an ontology-based Graph-RAG pipeline for contextual extraction and a Multi-LLM Gateway for dynamic runtime switching.
Challenges we ran into
The biggest challenge was ensuring the physical safety of AI-generated commands. AI models can sometimes suggest actions that are physically impossible in the real world (e.g., resource contention or hardware limitations). We solved this by developing an "Atoms > Bits" philosophy. We built a ConflictResolver that ensures physical constraints always override software instructions, backed by a whitelist-based RuleGuard.
Accomplishments that we're proud of
Bridging AI and Reality: We successfully built a true closed-loop system. Instead of stopping at generating dashboards or text recommendations, ConverterOS actually dispatches automated, executable commands to real-world systems.
The "Atoms > Bits" Architecture: We are incredibly proud of our ConflictResolver safety layer. Designing an enforcement mechanism where physical constraints (Atoms) reliably override AI hallucinations or logical errors (Bits) was a significant engineering feat.
Dynamic Multi-LLM Gateway: We built a flexible and robust API gateway that allows for runtime switching between different LLM providers (including Amazon Nova) without needing to restart the server.
Actionable Enterprise UI: We successfully transformed typical text-heavy AI responses into a sleek, metric-centric B2B dashboard. By utilizing deterministic JSON outputs, our UI provides instant, quantifiable value (like cost savings and KPI badges) to operators.
What we learned
We learned how to effectively engineer Amazon Nova to produce highly structured, deterministic outputs. By refining our prompts, we directed the model to return concise, metric-centric JSON payloads for our UI instead of verbose text. We also deepened our understanding of integrating multimodal context (such as ambient environment states) into a unified decision pipeline.
What's next for ConverterOS
We plan to expand our Graph ontology using Amazon Neptune for production scale. We will also integrate more diverse operational domains, such as healthcare facility management and smart infrastructure, proving that ConverterOS can successfully orchestrate any fragmented physical network.
Log in or sign up for Devpost to join the conversation.