Gunveer Sandhu

Designed Android call and SMS protection distributed via Google Play. Implemented screening logic mindful of Android’s hard five-second call-screening deadline by moving provider IPC and non-critical work off the decision path, then using deterministic rule precedence, fail-safe handling, and a repeat-caller safeguard to keep outcomes predictable under platform constraints. Kept SMS/MMS content on-device with a ten-category classifier and explicit user overrides. Built reactive Kotlin/Jetpack Compose and Room state that recovers across process recreation, default-SMS role changes, and partial provider failures.

Built an end-to-end private multimodal reflection journal across Flutter (iOS/Android), TypeScript/Deno Edge Functions, and Supabase/PostgreSQL. Connected private journal state, authentication, subscriptions, push notifications, and five-language localization while moving across client, backend, and data concerns. Separated deterministic calculations from generated interpretation by building a Vedic astronomy engine for planetary positions/timing, then layering nine schema-constrained Gemini workflows over its results (text and vision). Engineered resilience to timeouts and malformed model outputs with workload-specific routing, regional fallbacks, bounded retries, and JSON recovery; instrumented token use and feature-level cost tradeoffs. Enforced secure data access via Row-Level Security and made RevenueCat webhooks idempotent for repeated events; implemented versioned migrations, automated client/backend tests, and separate staging/production CI for controlled releases.

Owned a private on-device writing autocomplete for macOS. Designed a privacy-first architecture that avoids screen recording and cloud inference by leveraging app-specific Accessibility integrations to keep the product fast and lightweight on 8 GB Macs. Built typed adapters to separate host-specific caret/text access from context selection and autocomplete behavior, and revalidated live draft text and insertion positions before rendering or accepting suggestions. Created Python/MLX training tooling for raw-text completion and adapted OLMo 2 1B using LoRA across 16 layers with cursor-position examples matched to production prompts, including controlled evaluations against an unadapted quantized baseline. Optimized inference latency using KV-cache reuse, idle prewarming, one-ahead generation, and Swift Concurrency cancellation for superseded requests. Delivered model distribution with Developer ID signing, notarization, signed updates, and automated release checks plus content-free diagn