Product
Content Engine
A living, breathing content engine for sports: AI agents that build promotions and content across every channel, grounded in real fixtures and live data.
What is it?
A content and campaign platform built from eight specialist agents working together in one conversation: research and entity resolution find and ground the right content, promo creation turns it into a structured campaign brief, and the copy agents produce ready-to-send email, social and push variants, exported with a full audit trail.
What does it do?
- -Resolves ambiguous queries to the right entity before answering, never guesses
- -Grounds every campaign in real, live fixture data rather than assumed events
- -Produces email, social and push copy from the same picked content, no re-explaining the brief per channel
- -Every export is a tracked, tenant-isolated record with a full audit trail
Why does it matter?
We're an agentic-first development agency and technology business. Content Engine is that approach applied to iGaming content production: research, briefing and copy production as one connected workflow instead of separate tools.
Who is it for?
Marketing, Engineering
How it works
A chat-first interface backed by eight specialist agents (research, entity mapping, promo creation, and email/social/push copy, plus export), each grounded in the platform's canonical entity and fixture data rather than model recall alone.
Components
News Bulletin - Trending Topics Agent
Gives marketing teams a live read on what's actually happening across the sporting world right now, without manually scanning news and social all day.
Watches over 30 news outlets and, the moment new articles land, understands what's actually trending and turns it into a news bulletin ready for marketing teams to use in player communication.
- Autonomous and continuous, not scheduled: it fires off the real ingestion event, not a fixed hourly check.
- Self-checking before it ever reaches a person: every citation and headline is verified against the source articles before it writes, deterministically, no model grading its own homework.
- Never publishes on a failed check: if self-review fails, the write is blocked and the team is alerted, not shipped with a shaky citation.
- Cheap enough to run continuously: a lightweight model handles synthesis, chosen deliberately over a heavier one for a task that fires this often.
How it works: Claude Haiku, via a dedicated Gateway budget, triggered directly by the content-ingestion event stream. Runs a five-step pipeline: Plan (pull trending entities and the latest content), Synthesis (Haiku drafts the bulletin), Self-Review (deterministic citation and headline checks), Write (idempotent insert), Announce (fires a pointer event for downstream systems).
Research Agent
Resolves "who or what is this actually about" from a single ambiguous query, then pulls the real content behind it, without a person searching multiple systems.
An agent that resolves the specific player, team, or competition a query means, asking a clarifying question when a name is genuinely ambiguous rather than guessing, then streams the real articles, stats and picks behind that resolved entity.
- Refuses to guess on ambiguous entities: asks one clarifying question instead of surfacing the wrong sport's data with false confidence.
- Fast even cold: a two-tier cache serves the first request from a roughly 300-entity hot set in about 50 milliseconds, instead of a 9.8-second full-table fetch.
- Self-healing: a fallback point-query catches long-tail entities that aren't in the cache yet, instead of returning a false no-match.
- One resolved entity feeds three different content types, articles, stats, and picks, without re-resolving each time.
How it works: A tiered-cache resolver backed by the platform's canonical entity tables, paired with two streaming tools: one for articles and stats, one for fixture-based picks. Both source facts exclusively from verified data, never model recall.
Entity Mapping Agent
Turns a raw news article into correctly tagged, deduplicated entities the rest of the platform can query, the invisible work behind every "who's this about" answer.
A two-flow pipeline that runs on every ingested article: first extracts every named entity with its type, sport and confidence, then, for anything that doesn't match an existing record, makes a second-pass judgment call on whether it's a known alias or a genuinely new entity.
- Two models, two jobs: a fast, cheap model for high-volume extraction, a second model reserved for the harder disambiguation decisions.
- Never blocks ingestion: if entity resolution fails on one article, the article still ingests, it just stays unlinked rather than failing the whole pipeline.
- Grows its own reference data: every confirmed new entity or alias gets written back, so future articles resolve faster.
- Real scale today: 102,035 canonical entities and 10,954 known aliases, the entire set held in memory and refreshed every 15 minutes.
How it works: Flow 1: a fast, cheap model extracts named entities from article text via the AI Gateway. Flow 2: for unmatched surface forms, Claude Haiku judges each against up to 20 candidate matches and decides alias versus new entity, batched to stay well inside context limits.
Real 6-hour window — 100 articles ingested, entities extracted from each
Intelligent Promo Creation Agent
Turns a campaign idea into a complete, structured brief, grounded in real fixtures and the operator's own promo mechanics and brand guidelines, all through an interactive chat.
A conversational agent that captures the seven things every campaign needs, brand, focus, market, event, output format, mechanic, campaign name, one question at a time, picking up whatever's already implied in what you typed and only asking for what's missing. Underneath, it reads that operator's own live brand config for the bonus mechanics they're actually allowed to offer, and resolves the campaign's focus, a team or player, to their real next fixture: opponent, competition, kickoff, pulled from live fixture data rather than assumed.
- No blank-form paralysis: it asks one focused question at a time instead of presenting seven empty fields.
- Never offers a mechanic the operator can't run: available mechanics are read live from that brand's own config, not a generic list.
- The event is never assumed: focus resolves to the entity's actual next fixture, opponent, competition and kickoff time, pulled from live fixture data.
- Nothing forgotten: none of the seven slots can be skipped silently, every campaign starts from a complete brief.
- Edits mid-flow: change your mind on the market halfway through and it updates, not restart from scratch.
How it works: Claude Sonnet, called once per question in the flow, resolving the focus entity against the platform's canonical entity and fixture data and pulling that operator's live brand config for its approved mechanics as the conversation goes. Each call returns the current captured state plus the single next question to ask, rendered as a filled-in card with quick-pick options. A confirmed brief, entity, fixture and mechanic all resolved, hands straight to the content-generation agents.
Email Copy Agent
Gets a marketer from "I need an email for this" to three ready-to-send subject lines and three body options, without writing a first draft themselves.
An agent that takes whatever content the marketer already picked (an entity's stats, an offer, a confirmed angle) and streams three subject line options followed by three body copy options, in the brand's own voice, sourced and cited.
- Streams as it writes: options appear within about half a second, not after a long wait.
- Brand rules aren't a checklist a person has to remember: prohibited terms, tone, and channel rules are compiled in before a single word generates.
- Forward-looking language for fixtures that haven't happened yet is an enforced rule, not left to chance.
- Every option is picked from, not accepted blind: the marketer chooses, doesn't just approve.
How it works: Claude, streaming through the same chat interface as everything else in the platform, with that tenant's brand voice and content rules compiled in before generation starts. Designed for a sub-second time to first token.
Push Copy Agent
Gives marketers three ready push notification options, sized to the least forgiving format in the stack.
An agent that streams three numbered push notification options straight from picked content, with brand voice and mechanic guidance baked in.
- Same picked-content input as email and social: no re-explaining the campaign for a third time.
- Mechanic guidance (what the CTA should say for a price boost versus a free bet) comes from config, not guesswork.
- The responsible-gambling tagline logic is conditional and automatic, matching the same rule set as every other channel.
- No hardcoded regulatory language anywhere in the prompt, everything routes through brand config so it stays current as rules change.
How it works: Claude, same streaming mechanism as email and social, tuned for the tighter character constraints of a push notification.
Export Agent
Gets a finished asset out of the chat and into a usable, tracked record, without a separate tool.
An agent that, the moment a marketer confirms an asset, writes a permanent, tenant-isolated record: asset type, content snapshot, export format, campaign name, and timestamp.
- No LLM round-trip: it's instant and deterministic, nothing to wait for or get wrong.
- A queryable history: every export is a searchable database row, not a file that gets lost.
- Tenant isolation enforced at the database level, not by convention, one team can never see another's exports.
- The same action that finishes the campaign creates the audit trail, no separate step to remember.
How it works: A direct, authenticated write to the platform's asset records table, with row-level security enforcing tenant isolation. No model call in the loop, purely app-layer.
From “who’s playing tonight” to a scheduled campaign, in one chat
Discover the entity, build the promo, schedule it against the actual fixture. No separate tools, no re-typing the brief three times.
01: Discover
Say who it's about. It finds the fixture.
Type a player, team, or the vaguest question you've got, and it resolves against real fixture and squad data. No separate search, no picking the right database first.
Real example — canonical entity IDs, resolution state unchanged
Typed: “Haaland”
Auto-resolved, locked in:
Typed: “Silva”
Which Silva did you mean?
02: Build
The brief fills in as you talk.
Market, mechanic, output format, all captured one question at a time and grounded in that operator's own approved promo mechanics, not a generic template.
Real example — a resolved run, brand and mechanic unchanged
1. Entity resolved
03: Schedule
Pinned to the actual kickoff, not a guess.
Every scheduled promo is tied to a real fixture. No fixture, no schedule, the platform won't let a campaign go out ungrounded.
Real fixture — Champions League, tonight
04: Proof
Then it writes the campaign itself.
Email, social, and push copy generated from that same resolved brief, in the brand’s own voice.
Real example — a resolved run, brand unchanged
Subject
▍
Preview text
Body
Please gamble responsibly. 18+ only. For help and support visit BeGambleAware.org. T&Cs apply.
Social
Real example — a resolved run, brand unchanged
▍
18+ | BeGambleAware.org | T&Cs apply
▍
18+ | BeGambleAware.org | T&Cs apply
Push
Real example — a resolved run, brand unchanged
Social Copy Agent
Does the same job as the email copy agent, but shaped correctly for whichever platforms the campaign is actually going out on.
An agent that streams a section per picked platform, three subject and three body options each, with hashtag suggestions and CTA shapes pulled from that brand's own configuration, never hardcoded.
How it works: Claude, same streaming mechanism as email copy, with a platform-aware prompt block driven by that brand's channel configuration.