Case study
Dashboard IA
AI agent observability, finally readable.
Key facts
- Client
- Confidentiel — scale-up SaaS
- Role
- Lead developer & product designer
- Year
- 2026
- Duration
- 6 months
- Stack
- ReactSupabaseRedux ToolkitAPIs IA
01
Context and problem
A SaaS scale-up orchestrates dozens of AI agents automating support and operations for its customers. Every agent emits a continuous stream of events — decisions, tool calls, costs, errors — that teams were steering blind, between raw logs and spreadsheets.
The problem
Three simple questions had no answer: what are the agents doing right now, how much do they cost, and where do they fail? Customers spotted incidents before the team did, and every investigation took hours.
02
The solution
We designed a realtime control room: a fleet view aggregating each agent's state, replayable decision timelines and cost-drift alerts. The interface is built on a restrained design system made for information density.
Key features
Realtime fleet view
Each agent's status, current tasks and load, refreshed every second through Supabase Realtime.
Replayable timelines
Every agent decision is traced and can be replayed step by step to understand a failure in minutes.
Drift alerts
Per-agent cost and latency thresholds, notified before the bill runs away.
Emergency controls
Pause, restart or downgrade a model in one click, without redeploying.
03
Gallery
A look at the delivered screens — from the main journey to interface details.
04
Results and numbers
What the project changed, measured over time.
−68% investigation time
Replayable timelines replace hours of log digging.
−31% inference spend
Drift alerts stop token waste before it reaches the invoice.
92% of incidents caught before customers
The fleet view surfaces every failure on its first occurrence.
+54% internal adoption in three months
Support teams now steer their agents on their own.
Let's work together
A product to launch, a platform to harden, a team to reinforce? Let's talk.