SentinelFlow

Inspiration The global logistics industry is currently plagued by "Reconciliation Gaps"—the operational friction between physical shipment anomalies and the digital systems tracking them. While studying Software Engineering at ENSEM, I identified that identifying a single delay and notifying stakeholders currently requires manual cross-referencing across fragmented platforms. This inefficiency leads to an estimated annual loss of over $4B globally.

SentinelFlow was born from the desire to transform this reactive, labor-intensive process into a proactive, agentic workflow. I wanted to build a system that doesn't just display data, but "thinks" through disruptions to resolve them autonomously.

How I Built It The architecture of SentinelFlow is designed for enterprise-grade reliability and transparency:

Agentic AI Core: Built using LangChain to orchestrate multi-step "Thought-Action" cycles. The system uses Llama 3 (via Groq) to determine when to query databases and how to structure resolution drafts.

System Gateway: A high-concurrency FastAPI backend that manages real-time logistics data streams.

Enterprise Interface: Developed with Next.js 15, featuring a live Decision Graph to visualize the AI's internal reasoning process.

Infrastructure: The entire stack is containerized with Docker Compose to ensure seamless deployment and environment parity.

Challenges Faced Inference Latency: Agentic reasoning requires multiple LLM calls. To ensure real-time responsiveness, I integrated Groq LPU™ inference, achieving sub-second token generation speeds.

The Trust Gap: Autonomous AI can be unpredictable. I solved this by implementing a Human-in-the-Loop (HITL) protocol. The AI identifies and drafts solutions, but requires explicit human authorization for high-stakes actions, ensuring 100% data integrity.

State Management: Handling complex, multi-step reasoning without infinite loops required a robust state management layer within LangChain to track the agent's history and current objectives.

What I Learned Through this development process, I gained deep insights into the difference between simple RAG (Retrieval-Augmented Generation) and Agentic Orchestration. I learned that for AI to be useful in an enterprise setting, Observability is just as important as accuracy. By visualizing the reasoning process through a Decision Graph, the system moves from a "black box" to a transparent, auditable tool.