Naboria RMA — Technology Stack

An omnichannel campaign management platform with ML budget optimization, geo lift measurement, and real-time syncing across Meta, Google, and Naboria Ads.

Back to Stack Overview

Campaign Management Flow

How campaigns are created, synced to ad platforms, and monitored in real time.

Multi-Channel Campaign Sync

graph LR subgraph API["Go API Server"] CAMP["Campaign\nManager"] SYNC["Channel\nSync Engine"] ALERTS["Alert\nMonitor"] AUTH["JWT Auth\n+ Multi-Tenant"] end PG[("PostgreSQL\nCampaigns, Users,\nOrganizations")] subgraph Channels["Ad Platforms"] META["Meta\nMarketing API"] NADS["Naboria Ads\nInsights API"] GOOG["Google Ads"] end subgraph OLAP["ClickHouse"] PERF["Campaign\nPerformance"] CREAT["Creative\nPerformance"] FORE["Forecasts"] BUDG["Budget\nAllocations"] GEOR["Geo Lift\nResults"] end CAMP --> PG CAMP -->|push| SYNC SYNC -->|"every 5m: push campaigns"| META SYNC -->|"every 15m: pull insights"| META SYNC -->|"every 15m: pull insights"| NADS META -->|performance data| PERF NADS -->|performance data| PERF ALERTS -->|"every 15m: check 5 conditions"| PG

Push Active campaigns are pushed to Meta every 5 minutes — budget changes, status updates, and audience targeting. Pull Performance insights are pulled from Meta and Naboria Ads every 15 minutes into ClickHouse for real-time dashboards.

ML Pipeline

How raw performance data becomes budget recommendations, forecasts, and incrementality measurements.

ML Optimization & Measurement Flow

graph LR subgraph CH["ClickHouse"] PERF["Campaign\nPerformance\n(daily metrics)"] GEOP["Geo\nPerformance\n(by DMA)"] end subgraph Pipeline["Python / Dagster"] PROPH["Prophet\nForecasting\n(every 6h)"] HILL["Hill Curves +\nSLSQP Optimizer\n(daily)"] DID["DID + Permutation\nTest\n(every 6h)"] ATTR["Time-Decay\nAttribution\n(on demand)"] end subgraph Output["ClickHouse Output Tables"] FORE["campaign_forecasts\n(7-day predictions)"] BUDG["budget_recommendations\n(channel allocations)"] LIFT["geo_lift_results\n(lift, p-value, CIs)"] ATTRR["attribution_reports\n(ROAS, CPA, marginal ROI)"] end SERVER["Go API Server"] DASHBOARD["Next.js\nDashboard"] PERF --> PROPH PERF --> HILL GEOP --> DID PERF --> ATTR PROPH --> FORE HILL --> BUDG DID --> LIFT ATTR --> ATTRR FORE --> SERVER BUDG --> SERVER LIFT --> SERVER ATTRR --> SERVER SERVER --> DASHBOARD

ML Four distinct ML pipelines run on different schedules — forecasting and geo lift every 6 hours, budget reallocation daily, and attribution on demand. All outputs land in ClickHouse and are served through the Go API.

Campaign Alerting

Five automated conditions continuously monitor campaign health and flag anomalies.

Alert Monitoring Flow

graph LR subgraph Scheduler["Background Scheduler"] CHECK["Alert Checker\n(every 15m)"] end subgraph Conditions["5 Alert Conditions"] C1["Pacing Risk\n(projected spend > 110%\nof remaining budget)"] C2["Budget Low\n(spent > 85%\nof lifetime budget)"] C3["ROAS Drop\n(yesterday ROAS <\n7-day avg × 0.7)"] C4["Zero Spend\n($0 for 2+ of\nlast 3 days)"] C5["Spend Anomaly\n(yesterday > 14-day\navg × 2.5)"] end PG[("PostgreSQL\nAlert Records")] subgraph Notify["Notifications"] SLACK["Slack"] EMAIL["Email"] HOOK["Webhooks"] end DASH["Dashboard\nAlert Badge"] CHECK --> C1 CHECK --> C2 CHECK --> C3 CHECK --> C4 CHECK --> C5 C1 --> PG C2 --> PG C3 --> PG C4 --> PG C5 --> PG PG --> SLACK PG --> EMAIL PG --> HOOK PG --> DASH

Alert The scheduler checks all active campaigns against 5 conditions every 15 minutes. When triggered, alerts are persisted to PostgreSQL and dispatched to configured notification channels — Slack, email, or custom webhooks.

Geo Lift & Experimentation

Measure true incremental impact with geo-based Difference-in-Differences and A/B experiments.

Incrementality Measurement

graph LR subgraph Setup["Experiment Setup"] DMA["DMA Regions\n(treatment vs control)"] AB["A/B Variants\n(audience splits)"] end subgraph Data["Performance Data"] GEOP["Geo Performance\n(daily, per DMA)"] CAMP["Campaign Metrics\n(per variant)"] end subgraph Analysis["Statistical Analysis"] DID["Difference-in-\nDifferences"] PERM["Permutation Test\n(p-value)"] ZTEST["Two-Proportion\nZ-Test"] end subgraph Results["Results"] LIFT["Incremental Lift\n+ Confidence Intervals"] WINNER["Winning Variant\n+ Significance"] end DASH["Dashboard\n(Leaflet maps +\nresult tables)"] DMA --> GEOP AB --> CAMP GEOP --> DID DID --> PERM PERM --> LIFT CAMP --> ZTEST ZTEST --> WINNER LIFT --> DASH WINNER --> DASH

Components

Every piece of the Naboria RMA stack and what it does.

API Server

Go / Chi

A fast REST API handling campaign CRUD, multi-tenant auth, channel syncing, and background scheduling — all in a single binary.

Frontend

Next.js + React

A TypeScript SPA with Tailwind CSS, Leaflet maps for geo lift visualization, and a JWT-aware API client with automatic 401 redirect.

Relational DB

PostgreSQL

Stores the campaign lifecycle — campaigns, budgets, creatives, audiences, experiments, alerts, and multi-tenant organization hierarchy.

Analytics DB

ClickHouse

Time-series performance storage using SummingMergeTree and ReplacingMergeTree — campaign metrics, forecasts, budget recommendations, and geo lift results.

Cache

Redis

Session cache and rate limiting for the API layer, keeping hot data fast and reducing database load.

Forecasting

Prophet

Generates 7-day performance forecasts every 6 hours — used for pacing alerts and budget planning.

Budget Optimization

Hill Curves + SLSQP

Fits diminishing-returns curves to historical spend and uses constrained optimization to allocate budget where marginal return is highest.

Incrementality

DID + Permutation Tests

Geo-based Difference-in-Differences with permutation tests for statistical rigor — proving true incremental lift across treatment and control markets.

Experimentation

A/B Testing

Two-proportion z-test for comparing variant performance with statistical significance — applied to creative, audience, and channel experiments.

Attribution

Time-Decay WLS

Weighted least squares attribution model with time-decay weighting — allocating credit across channels based on recency and contribution.

Orchestration

Dagster

Schedules all ML pipelines — forecasting, budget allocation, geo lift, and attribution — with job-level observability and retry logic.

Reverse Proxy

Caddy

Path-based routing — /api and /auth to the Go server, everything else to Next.js — with automatic TLS via Let's Encrypt.

Channel Integration

Meta Marketing API

Full bi-directional sync via Graph API v21.0 — push campaigns and budgets, pull daily insights and creative-level performance.