Recommendation Intelligence for Social-Commerce Marketplace

---

Designing, specifying, and validating a multi-engine recommendation system across a three-sided marketplace — delivered as a complete, evaluable methodology with an evidence-based readiness assessment.

Recommendation Systems · ML Product Management · Independent Model Validation · Decision Intelligence

The client was an early-stage social-commerce platform combining influencer-led content discovery with an integrated product marketplace — a three-sided marketplace where consumers discover products through creator content, promoters and influencers earn commission on what they drive, and brands run campaigns against that audience. The platform had already shipped twenty modules, including a dynamic interest-scoring engine, an influencer analytics funnel, and an onboarding interest graph. What it lacked was the decision layer that determines what each user sees and how much each contribution is worth, across three distinct surfaces: consumer recommendations, opportunity matching for promoters, and campaign–promoter matching for brands. Cristo Labs was engaged to own that layer end to end.

The engine had to be relevant from launch, evolve as the platform grew, and stay commercially trustworthy — under real data constraints that ruled out simply training models on day one. A pre-engagement audit confirmed the binding gaps: the order model carried no return, refund, or cancellation state, so no negative commerce outcome could be labeled; instrumentation captured only a handful of coarse actions, with no impressions, positions, skips, or explicit feedback, meaning no exposure data and therefore no negatives to learn from; there was no video-playback telemetry, no persisted promoter or influencer state over time, no campaign entity for brands to recruit against, and no content lineage for reposts. Layered on top were non-negotiable commercial boundaries — commission must never influence what a consumer is shown, brands must never be ranked as candidates, and every score reaching a user or brand had to be explainable and reproducible or it would not ship.

Three workstreams ran in parallel against one governing proposal — platform and product, engineering, and data science — with a single boundary rule keeping them clean: data science owns the algorithms, and engineering owns everything that makes them run. Models, features, labels, scoring, calibration, confidence, exploration policy, cold-start strategy, and attribution sat with Cristo Labs; the serving spine, feature store, eligibility gates, and enforcement sat with engineering. Two rules resolved the rest — a threshold tuned from data is an algorithm, while a threshold set by policy or law is not; and deciding a rule is never the same as enforcing it. That discipline is what kept marketplace ranking from ever coupling to payout policy.

Rather than promise mature models the data could not yet support, the work was structured into waves gated on data conditions, not calendar dates. Wave 1 delivered a transparent baseline: interest scoring with recency decay, rule/content/graph/vector retrieval, gradient-boosted or compact rankers, Bayesian-shrunk promoter scores, deterministic reranking, and rule-based fraud controls. Wave 2 reserves production ML — embeddings, two-tower retrieval, an approximate-nearest-neighbour index, a multi-task feed model, a two-stage product model, and probability calibration. Wave 3 reserves mature optimisation — mixture-of-experts and sequence models, distillation, learned calibration, graph-based fraud, principled exploration, and causal uplift. Each of nine model families was specified in both an initial form that runs on data available at launch and a mature form that requires accumulated volume, making the plan a data-readiness statement rather than an unresolved design choice.

Wave 1 produced twelve specification, methodology, and evaluation artefacts, each written to be implementable, evaluable, and acceptance-tested. These covered feature definitions and semantics across user, content, promoter, product, campaign, context, and interaction families; label construction for every prediction target, positive and negative, with latency bounded by the return window; a recency-decayed interest and profile model that reproduces the platform's existing score exactly when decay is disabled; retrieval methodology across rule, content, graph, and vector sources; ranking methodology and objective design with an explicit cap on the commerce term and commission excluded from consumer-facing objectives; promoter scoring with Bayesian shrinkage, confidence bands, and normalisation by category, region, account size, and format; opportunity and campaign methodology with expected-earnings and expected-reach ranges; social proof, cold-start and exploration strategy, attribution methodology, a reason taxonomy and confidence methodology, and an offline evaluation protocol using walk-forward evaluation.

The methodology built commercial and ethical guardrails into the design rather than bolting them on: a commission guardrail, enforced by acceptance test, so a low-relevance, high-payout item can never be elevated on a consumer surface; brand exclusion from the campaign engine; point-in-time correctness, so promoter standing is read as at the time of each exposure and cannot predict its own earlier sales; and a rule that any output reaching a user or brand carries a reason code and version, so a score that cannot be explained and reproduced is not delivered. Its most distinctive output was a candid one — a precise, evidenced statement of what could and could not yet be trained, with every blocking dependency named and assigned to an owner, and a binding constraint that no numeric finding could be reported from test or synthetic data. This converted hidden data debt into a managed programme plan and set the earliest credible activation date for each advanced model family before any production code depended on it.

Wave 1 was delivered and accepted. The client received a complete, implementable recommendation methodology across three engines and nine model families; a launch-ready, transparent baseline that extends the platform's existing interest engine and degrades safely back to it, so no failure is ever user-visible; an offline evaluation protocol and a per-family readiness assessment stating, with evidence, what is ready to train and what is not; and a dependency and instrumentation roadmap the platform and engineering teams can execute against. The result is a scalable personalization foundation designed to launch now and grow with the platform, without rebuilding the recommendation system at each stage of adoption. Waves 2 and 3 are part of the target architecture and activate only as data-readiness gates are met, family by family — Wave 2 as impression, watch, commerce, and return data reach measured volume, and Wave 3 once Wave 2 is operating in production with a body of repeat experiments behind it.

The engagement demonstrates Cristo Labs' core capabilities in one place: recommendation systems; ML product management from problem formulation through methodology, validation, documentation, and handover; independent, reproducible model methodology; decision intelligence and scoring governance; and enterprise delivery discipline, with defined scope, acceptance criteria, responsibility assignment, and change control.