Platform
How Predictively works.
A controlled pipeline from verified source data to quality-scored, structured intelligence briefs. Every step is deterministic, every output is schema-validated.
Pipeline
From raw data to decision-ready output.
Data ingestion
Monthly operator filings from the Ontario Gaming Commission are ingested through a deterministic pipeline with period-over-period reconciliation. No manual entry. No editorial interpolation at the source layer.
Source pack construction
The ingested data is structured into a dense source pack covering monthly timelines, quarter-to-date comparables, trend signals, decision signals, and evidence bullets. The source pack is the authoritative input to the generation step.
Schema-constrained generation
The intelligence engine generates a structured JSON brief against a fixed schema. Instructions require every claim to be anchored to the source pack. Invented figures, unsupported causal claims, and generic language are treated as quality failures.
Quality scoring and repair
The generated brief is scored on specificity, commercial utility, credibility, actionability, and non-generic language. Briefs below the quality threshold are repaired or regenerated against the weak sections before release.
Structured output
The validated payload is rendered into the four-module briefing format: executive posture, opportunity radar, risk mapping, and recommendations. The same source lineage powers both the intelligence brief and the data visualisation layer.
Design decisions
Why the product is built this way.
A fixed output schema means leadership learns the format once, then reads it faster every cycle. Structural consistency also makes quality scoring tractable.
A chat interface shifts interpretive burden back to the user. Predictively is built to do the interpretation work, not to provide a raw surface for users to do it themselves.
Every claim must trace back to a verified data point in the source pack. This makes the output auditable and prevents the hallucination patterns that render unconstrained AI research unusable in professional settings.
When a fiscal quarter is incomplete, the platform compares the first N months of the current quarter against the same N months of the prior-year equivalent quarter. Partial-quarter-against-full-quarter comparisons are not used.
Operator strategy, investor brief, and compliance watch framings adjust how the intelligence is positioned without changing the underlying evidence base.