Project Antigravity
Most recently I architected and built a multi-tenant-ready customer-acquisition and data platform — a live acquisition front-end, a probabilistic identity-resolution data engine, an autonomous real-time bidding engine, and a built data-productization pipeline — solo, end to end.
The problem
Customer acquisition in a high-volume, compliance-heavy vertical normally takes three teams: a conversion-optimized traffic front-end, a data system that resolves and deduplicates identities at scale, and a media-buying operation that bids on ad inventory profitably in real time. I set out to architect all three as one coherent, largely autonomous system.
The acquisition front-end
The platform's acquisition front-end is a Next.js 16 / TypeScript (strict) paid-traffic funnel across multiple verticals, integrated with the Boberdoo ping-post lead marketplace, with production TCPA-compliance infrastructure (Jornaya LeadiD + TrustedForm + a Supabase consent ledger): 21 routes, zero type errors.
The acquisition front-end of the platform, TallyRate.com, is live.
Visit TallyRate.comThe identity-resolution data engine
The platform's identity-resolution data engine is a batch-ETL performing Splink v4 Fellegi-Sunter probabilistic record linkage (EM-trained parameters, connected-component clustering) into an Apache Iceberg → Snowflake hub-and-spoke warehouse, with tagged releases v2.0–v2.4.1, a 77-test suite, and an O(n²)→O(n·k) blocking optimization.
A blocking optimization reduced candidate-pair generation from O(n²) to O(n·k) via composite geographic + name-locality keys.
Consent governance runs as a daily revoke-scan producing tombstone manifests, with warehouse secure views filtering revoked records — regulatory compliance enforced at the data layer, not in application code.
The autonomous bidding engine
The platform's autonomous bidding engine drives real-time bid decisions inside a 150ms budget from a LightGBM → ONNX CTR-prediction model (14-feature), with recursion-based yield optimization and Meta/Google/OpenRTB adapters — and it is air-gapped: zero credentials, anonymized telemetry only.
I air-gapped the autonomous engine on purpose: the highest-risk component is the one with the least access — credentials and PII are structurally out of its reach.
The data-productization pipeline
The platform's capture→resolve→warehouse→productize path to S3 + Snowflake is built and wired — turning every interaction into deduplicated, consent-governed records ready to package as industry data.
Status, stated plainly
The platform is architecturally complete; the front-end is live and the bidding engine has run on real ad spend. The ML model ships trained on industry-benchmark priors and retrains on real traffic — a deliberate cold-start design.