Operators asked for two things: fill every community table with the right mix of guests, and keep the outreach personal even at scale. We answered with a pair of cooperating microservices riding on MongoDB vector search, orchestrated through an API gateway, and fueled by OpenAI prompts fine-tuned for Cuculi's tone.
Blueprint Snapshot
- 1Vectorized Personas: embeddings crafted from dining history, spend, neighborhoods, and social goals.
- 2Event Fingerprints: cuisine, vibe, atmosphere, and operations metadata collapsed into a cosine-friendly vector.
- 3API Gateway: shared envelope for auth, throttling, and fan-out to the recommendation microservices.
- 4AI Copy Studio: LLM prompts tuned for warmth, urgency, and segmentation-specific CTAs.
Vector-First Architecture
Every interaction starts with vectors. MongoDB Atlas hosts twin collections: diner_embeddings and experience_embeddings. Each record stores the dense vector, JSON metadata, and lightweight counters we use for scoring. Incoming payloads hydrate a VectorContext object that pushes through the API gateway and into the respective microservice.
Diner Context Pack
- • Embedding derived from recency, spend, and interest taxonomy
- • Journey markers: onboarding stage, churn risk, loyalty tier
- • Social graph hints: crew affiliation, preferred group size
Experience Context Pack
- • Event fingerprint embedding with cuisine, vibe, price, host style
- • Capacity & pacing metadata for operations
- • Sponsorship slots, upsell hooks, and dietary coverage
Service Responsibilities
The engine is split into two services that share a vector utility library but maintain independent scaling policies. The User Recommendation Service focuses on active diners and high-value repeat guests. The Event Targeting Service scouts for audience pockets across the entire member base, prioritizing churn recovery and segment coverage.
User → Event Flow
- 1. Pull diner vector and nearest event candidates via Atlas vector search.
- 2. Score compatibility with contextual boosts (neighborhood, spend alignment).
- 3. Draft AI messages tuned to the diner's tone profile and risk status.
- 4. Return ranked list with justification tags for operator review.
Event → User Flow
- 1. Fetch event fingerprint and target compatibility threshold.
- 2. Query diners by cosine similarity with segment and churn filters.
- 3. Tailor AI outreach for re-engagement, upsell, or newcomer welcome.
- 4. Produce send-ready messaging with reasons to believe.
Lifecycle of a Recommendation
- Capture: Product surfaces send a request with the diner or event context, including lightweight analytics on recent sessions.
- Vector Search: Atlas executes a cosine search with dynamic filters, capping latency under 120ms per query.
- Scoring: We blend the similarity score with engagement heuristics and operator business rules.
- Messaging: OpenAI generates two variants — conversational SMS and polished email — both logged for experimentation.
- Feedback: RSVP actions stream back into the embeddings pipeline to keep future recommendations fresh.
Experiments and Instrumentation
Every campaign run becomes a data point. We log match_reason and targeting_reason tags for post-mortem analysis, track CTA performance by message variant, and pipe anonymized telemetry into Looker dashboards. The Gemini and Claude prototypes linked above shaped our copy strategy before writing a line of production code.
- • Campaign Studio: Operators preview AI copy, tweak CTA emphasis, and lock sends.
- • Feedback Loop: RSVP outcomes stream back into the vector store nightly.
- • Guardrails: Claude narratives stress-test tone, while Gemini scenarios validate event coverage.
30-Day Launch Playbook
Want to ship this stack in a month? Here's the sprint breakdown we used.
Week 1-2: Foundations
- • Instrument profile data exports and define embedding schema.
- • Stand up API gateway with mocked service responses.
- • Build prompt library for outreach templates.
Week 3: Vertical Slice
- • Implement vector search queries and scoring formulas.
- • Connect OpenAI for personalized copywriting.
- • Ship the operator dashboard (see the flat demo above).
Week 4: Launch & Iterate
- • Run closed beta with high-touch hosts and concierge team.
- • Measure conversion lift, calibrate message tone, and expand cohorts.
- • Activate retention automations using targeting reasons.
Evergreen
- • Quarterly prompt refresh with Claude narrative probes.
- • Embed health dashboards into ops rituals.
- • Refresh embeddings monthly with the latest RSVP signals.
Ready to give the architecture a spin? Start with the Cuculi AI Microservices Demo, then layer in live data sources when you're ready to launch. Every component here was shaped by real operator interviews and real diner feedback.