Software Engineer & Product Lead

The entire technical team for a production AI-first productivity platform, and the bridge between the founder's product vision and technical reality. Owns full-cycle engineering — AI pipelines, cloud infrastructure, App Store compliance, CI/CD — and every product decision that goes with it, converting unstructured emails, PDFs, and images into structured calendar events for consumers.

• 1Made the app 97% faster for users by replacing blind 15-minute email polling with custom mailhooks (Gmail push notifications), cutting average processing time from 20 minutes to 40 seconds — entirely my own initiative
• 2Cut cloud costs by 91% by building a junk filter that discards the ~90% of inbound attachment data (signature images, decorative graphics) that was useless noise before it ever reached the AI
• 3Rebuilt the entire app from no-code to full-code one month into the role, using AI dev tools to ship a working prototype in a single day that already outperformed a year of prior Adalo/Make.com work
• 4Engineered an AI email ingestion pipeline using OpenAI Responses API with strict JSON schema, native PDF input, and timezone-aware date resolution via Luxon, fully eliminating the temporal hallucination issues that plagued earlier LLM extraction
• 5Implemented AES-256-GCM field-level encryption with server-side DEK wrapping via GCP Secret Manager; built a serialized async queue to prevent iOS Keychain concurrency crashes caused by overlapping Riverpod providers
• 6Provisioned GCP infrastructure as code with Terraform (8 .tf files covering Pub/Sub, IAM, Secret Manager, API services) and authored three GitHub Actions CI/CD workflows for PR validation (format, lint, test) and dev/prod Android AAB builds
• 7Navigated an App Store Guideline 3.1.1 rejection and shipped a platform-split paywall: iOS uses App Store offer code redemption via RevenueCat; Android retains the approved in-app voucher flow
• 8Built multi-auth (Email/Password, Sign in with Apple, Google Sign-In) with platform-specific routing, anonymous-to-linked account flows, and branded transactional email via Gmail API
• 9Built an offline AI-extraction eval harness (immutable run records, deterministic scorer, multi-provider abstraction) so model and prompt changes never ship blind, then used it to run a real A/B on reasoning-effort levels — an evidence-based product call that cut p50 latency ~72% for a 3.4-point accuracy trade-off
• 10Instrumented the AI pipeline end-to-end and read 14 days of production logs to pinpoint the OpenAI call as ~96% of total task time — turning "the app feels slow" into a precise, defensible engineering target