The dashboard treadmill
Dashboards are answers to questions someone asked once. The business changes, the questions change, and the dashboards stay — so teams either stop looking at them or file another BI ticket and wait. Meanwhile the interesting events in the data (a segment quietly churning, a cost curve bending, a region outperforming) go unnoticed until someone stumbles into them, usually late.
ORACLE inverts the model. Instead of humans asking questions of static dashboards, specialized AI agents continuously watch the connected data, investigate what changed, and deliver findings — ranked by estimated business impact and strength of evidence — into a daily insight feed. When a finding raises a question, you ask it in plain language and the agents go investigate.
Before agents: an understanding of the data
Autonomous analysis on data nobody has defined is how you get confident nonsense. So ORACLE starts by building an understanding of the data, in two layers:
- Agents map the territory. Schema discovery walks the connected sources — starting with read-only PostgreSQL — and maps tables, relationships, and data-quality issues.
- Humans confirm the meaning. The business definitions that matter — what counts as revenue, what an active customer is — are confirmed by the team, not guessed by a model. Every later analysis builds on those confirmed metrics.
An orchestra, not a chatbot
A single do-everything agent is unaccountable — you can't budget it, bound it, or debug it. ORACLE uses an orchestrator that assigns work to dedicated agents, each with a narrow job and explicit limits on execution time and cost:
Discovery agents
Monitor sources for trends, anomalies, and changes across business segments.
Investigation agents
Compare periods, customer groups, and metrics to explore what may explain a change — observations clearly separated from hypotheses.
Forecasting agents
Project the metrics that move decisions, with confidence intervals rather than point promises.
Verification agents
Re-check analyses — queries, joins, definitions — before anything is presented.
Reporting agents
Turn verified findings into charts, summaries, and suggested next steps.
Trust is the product
An analytics platform that's occasionally, confidently wrong is worse than no platform. Most of ORACLE's design budget went into making findings verifiable rather than merely persuasive:
- Every finding shows its work — source references, the actual queries, data freshness, and caveats travel with the insight.
- Verification is a separate role. The agent that checks an analysis is never the agent that produced it.
- Observation ≠ hypothesis. "Enterprise churn rose 12% in March" and "possibly related to the February price change" are typed differently, displayed differently, and never blended into one confident sentence.
- Impact × evidence ranking. The feed is prioritized by estimated business impact and strength of evidence — so the top of the feed is worth a leadership minute.
Security posture: read-only until proven otherwise
ORACLE is designed for the questions a security review will ask. Access is read-only by default and every query runs sandboxed. Permissions are inherited — the platform can't show a user data they couldn't already see. Tenants are isolated, sensitive data is protected, every investigation leaves an audit history, and any write or external action requires explicit approval. This is the same posture we apply to every AI system we deliver: governance designed in, not retrofitted.
The MVP: deliberately narrow
The first release pairs one connector with the full trust loop, rather than many connectors with a shallow one: a PostgreSQL connector, schema discovery, human-confirmed metrics, discovery and verification agents, and a daily insight feed with charts and conversational follow-up. The foundation is modular — Next.js for the interface, Python/FastAPI services, Temporal for durable agent workflows that survive restarts mid-investigation, and PostgreSQL for platform metadata — so connectors and agent roles extend without re-architecting.
Design targets — first release
ORACLE is in active development; targets are the acceptance criteria of the MVP described above.
What the design process taught us
- Confirmed metrics are the contract. The single highest-leverage feature is the unglamorous one: humans signing off on what "revenue" means before any agent computes it.
- Budgets make agents governable. Time and cost limits per task turn "autonomous" from a risk into a line item.
- Durability beats cleverness. Real investigations take minutes to hours; workflow durability (Temporal) matters more than shaving seconds off a prompt.
- Separate the checker from the doer. Verification as an independent agent role caught more errors in design reviews than any prompt improvement we tried.
The one-line summary: ORACLE replaces "build me a dashboard" with "show me what changed, prove it, and take my follow-up questions" — with every answer carrying its evidence.