Moloads

# the idea

Own the buy side of programmatic advertising: accept a bid request, decide in milliseconds whether that impression is worth buying for a campaign, and close the loop all the way to a paid conversion — so advertisers pay for outcomes, not impressions.

# what I built

• OpenRTB 2.x bid endpoint with schema validation and a hard processing deadline (tmax).
• Bidding engine: banner-size matching, priority-weighted campaign selection, bid floors and an ML bid optimiser.
• Targeting on geo, device, OS, browser, time-of-day, app/site, audiences and first-party segments.
• Private marketplace deals — private-auction and preferred deals, attributable per deal and per supply route.
• Atomic per-impression budget reservation in Redis, so daily and total caps hold across every Cloud Run instance.
• Signed win notices, idempotent win recording, click tracker, impression pixel and server-to-server postback attribution.
• Fraud checks — IP/UA heuristics, per-IP rate limiting and blocklists; supply-chain validation; GDPR erasure.
• Four-role access model (super admin, account manager, agency, client) — margin fields stripped server-side for external roles.
• Report builder and time series across the full funnel: bids → wins → impressions → clicks → conversions → spend → revenue → ROI.
• Admin UI with advertisers, campaigns, creatives, offers, deals, rules, segments and monitoring.

# key decisions

• One FastAPI service instead of microservices — the latency budget is milliseconds, and network hops are the enemy.
• Deploys gated on a green CI run (700+ tests on SQLite and PostgreSQL, migrations and a live-bid smoke test), with automatic rollback on failed health checks.
• Security boundaries enforced in the API, not the UI — an agency literally cannot receive margin data.
• Redis for cross-instance state (budgets, frequency caps, rate limits); PostgreSQL as the system of record.

# stack

• Python
• FastAPI
• PostgreSQL
• Redis
• Alembic
• Cloud Run
• Cloud SQL
• Cloudflare
• GitHub Actions