Subcore Dashboard App

As AI agents move from demos into production, many teams that build and operate them can face a new problem: visibility. When an agent fails in the middle of a conversation, escalates to a human, or produces the wrong answer, we can’t identify why or even if that happened at all. Subcore was built to solve this. We designed and built the full-stack observability platform, giving AI teams a real-time window into every conversation, tool call, and decision the agents make.

Visit Website

A New Class of Infrastructure

AI agents don’t behave like traditional software, a conversation can involve dozens of events, like agent starting, an LLM call returning, a tool executing, a memory write, a human escalation, etc, all arriving in different ways. Subcore was designed around this reality from day one, building a purpose-designed event model covering the full agent lifecycle, storing and grouping every agent into sessions, and giving operators a coherent view of each conversation from start to finish.

Turning Raw Events into Actionable Intelligence

Logs tell you what happened and Subcore tells what it meant, with Google Gemini via the Vercel AI SDK to automatically generate summaries of every session with a concise request description, a resolution narrative and a readable status. The summaries are generated server-side with strict schemas, ensuring consistent, type-safe output that flows seamlessly into the dashboard and analytics. Teams can now review thousands of conversations at a glance, identify failure patterns and improve the events that caused them, without reading raw logs.

Real-TIme At Once

The people monitoring the events need to know what’s happening now, not after a page refresh. Subcore uses Supabase Realtime to push live updates to the dashboard the exact moment new events arrive or summaries are generated. We built a layered provider architecture, that separates concerns while delivering a seamless live experience. The result is that all the data updates in place without a full reload to every connected client.

Human-in-the-Loop

One of the most distinctive capabilities is support for human-agent handoffs. When an AI agent finds a conversation needing human attention, it fires an escalation requested event, where a human operator can claim the session, exchange messages and have control back to the agent. With the full HITL workflow: custom per-agent backend configuration, a connection-test API, claim and handback endpoints, and a conversation view that shows human and AI turns in a single thread.

Analytics

Subcore shows a picture of: total conversation volume, average session length, escalation rates, peak hours, and category-level breakdowns. The analytics, built and powered by Supabase RPCs, gives teams the data they need to understand their agents performance over time, spot regressions after deployments, and make clear and confident decisions about where to improve.

Final Results

Subcore transformed the reality of running AI agents in production into a clear, actionable dashboard. Teams using it can trace any conversation to its root cause in seconds, understand aggregate behavior at a glance, and use human help if necessary, all in real time. We delivered a platform that helps to monitor AI systems in the way they actually work, giving teams the visibility they need to make decisions in time and trust their systems more.

Metadata

Sector

Scope

Technologies