AI Application Engineer

Building production-grade AI for elite cricket is one of the most interesting engineering problems in sport.

Cricket Connect is an AI-native cricket intelligence platform working with several Vitality Blast county teams, a major franchise league, and a national team. We turn ball‑by‑ball data into pre‑match plans, in‑match decisions, and post‑match insight for coaches, analysts, and fans. Two products on a shared intelligence engine: CC AI (live match companion for fans) and CC Pro (professional analytics suite for coaches, analysts, and franchises).

The platform is live, the clients are professional cricket organisations, and the bar is rising fast. We're hiring the engineer who will help us raise it.

Cricket is one of the most data‑rich sports in the world — every delivery generates dozens of structured signals. Turning that volume of data into reliable, decision‑grade output for professional coaches is a meaningfully harder problem than building a typical LLM application:

• Numerical correctness is non‑negotiable. A coach reading a pre‑match report can spot a wrong number instantly. The system has to be right, not just plausible.
• The data is deep and the schema is rich. Ball‑by‑ball tables, innings tables, match tables, player tables — and the right answer often depends on joining them in exactly the right way.
• The output is structured. Pre‑match reports, matchup matrices, captaincy intelligence, breaking‑point analysis — every module is a contract between the data layer and the narrative layer, and that contract has to hold.

The engineering thesis we're building on is straightforward: the model should narrate structured data, not compose numerical claims from raw queries. The next phase of the platform is about deepening that discipline — precomputed views, canonical chapter payloads, validation layers between data and model, eval suites built from real cricket scenarios. That's where this role sits.

You'll work directly with me and our technical lead, owning the data and validation layer underneath both products. The first 90 days are concrete:

• Schema discipline. Document column semantics across the analytical tables (with COMMENT ON COLUMN and a maintained data dictionary), so every aggregation downstream is built on shared, explicit definitions.
• Precomputed analytical views. Build the canonical aggregations that power our most‑used modules — team phase‑state tables, defending‑band tables, chase‑state tables, matchup matrices — so query composition is centralised and consistent.
• Chapter payload contracts. Define the structured payload each module of the CC Pro match report returns: every number, every supporting innings, every denominator. The model narrates the payload. The numbers come from the data layer.
• Validation layer. Build the layer between data and model that enforces arithmetic consistency, scope sanity, denominator presence, and cross‑aggregate reconciliation — so every chapter ships with the same standard of rigour.
• Evals and regression tests. Establish the eval set so we can ship new modules with confidence and catch issues before users see them.

After that, you scale the same discipline across every new module we ship (player matchups, batting plans, bowling plans, captaincy intelligence) and set the engineering standard for whoever joins after you.

• LLM application layer. Heavy use of structured outputs, tool use, retrieval, evals. The model itself is not the bottleneck; the system around it is where the value is built.
• Database. Postgre
  
  SQL, ball‑by‑ball cricket data at scale (millions of rows), strong analytical query workloads.
• Engineering culture. Pragmatic, deliberate, evidence‑led. We precompute the view, validate the payload, narrate the result.

• Built production systems on top of LLMs where reliability matters and users will notice errors immediately. You treat "the model said something plausible but wrong" as a class of bug that needs systematic defence, not optimism.
• Strong SQL and data modelling. Comfortable designing schemas for analytical workloads, writing window functions and CTEs fluently, and deciding…