José Ignacio Feliú Fernández

Attribution Engine

Eight attribution models — first-touch, last-touch, last non-direct, linear, position-based, time decay, 50/50 split, and custom weighted — all computed at query time via Postgres RPCs. No pre-stored credits; switch models instantly without reprocessing. Identity stitching links anonymous visitors to known customers across devices. Synthetic touches are created for off-site conversions (e.g., GHL leads arriving via Meta) by matching lead metadata against prior session history.

Tracking SDK

A standalone JavaScript tracker distributed as an IIFE snippet and npm package. 30+ auto-capture modules — page views, form submissions, Stripe checkouts, Shopify purchases, GHL funnel forms, outbound links, video engagement, scroll depth, Web Vitals, and custom events. Events persist to localStorage and batch-send with exponential backoff. Bot filtering, AI agent detection, GDPR consent handling, and SPA navigation tracking out of the box.

Ads Explorer

Scrapes live competitor ads from Meta and TikTok — creatives, headlines, body copy, media. ML-powered scoring ranks ads by engagement potential. Extracts keyframes and audio from video ads. Feeds directly into the Ads Lab for variation generation.

Ads Lab

Multi-agent system for ad creation, built on Vercel AI’s ToolLoopAgent. Specialized agents — coordinator, research, psychology, copywriting, video scripting, compliance, and a critic that triggers retry loops for quality — orchestrated in a pipeline that produces ad copy, video scripts, and full creative briefs informed by competitor data from the Explorer and brand context from Notion.

Analytics & Reporting

Real-time dashboards with dual-period KPIs, funnel analysis, channel breakdowns, customer journey visualization, and AI-generated insights. Scheduled reports via email, PDF exports, and shareable public links. A public API with 25+ endpoints, rate limiting, and scoped API keys.

Space 3 Smartwatch Platform

Created the initial launcher and led development across 20+ applications for SoyMomo’s flagship children’s smartwatch. Built the chat app, phone dialer, phonebook, camera, gallery, keyboard, settings, step counter, reminders, friends, app store, and authentication manager — all running on constrained Android hardware with custom system-level behaviors.

Developed and integrated an on-device keyword spotting model using MFCC feature extraction, a CNN classifier, and SVDF layers, enabling streaming audio intelligence on the watch while meeting strict latency and power constraints. This work spearheaded the launch of a smartwatch generation that surpassed previous models and boosted SIM-card adoption.

Dangerous Content Detection — NLP Pipeline

University thesis project, deployed in production

Designed and implemented a multi-stage NLP moderation pipeline for SoyMomo’s Android tablets to detect cyberbullying, grooming, and explicit content in near real time. The system works in layers:

• OCR ingestion — replaced accessibility-based text extraction with screenshot-based OCR using Google ML Kit. Periodic capture, deduplication, local persistence, and cleanup workers.
• Dictionary models — weighted scoring for mood, cyberbullying, and grooming terms with regex patterns, slang expansion, and bilingual coverage.
• On-device classifier — evaluated RoBERTa, DeBERTa, BETO, and RoBERTuito; selected XLM-RoBERTa for multilingual performance. Fine-tuned on 20,000 curated examples, converted to ONNX, and quantized — reducing memory by 40% and increasing inference speed by 30%.
• Rust/JNI tokenizer — integrated Hugging Face tokenizers in Rust via JNI for performance-sensitive inference on Android.
• Cloud verification — batch processing with GPT-4o using JSONL batches and scheduled backend jobs for second-stage validation.
• Parent-facing alerts — real-time notifications with labels, screenshots, and detailed detection views.

Experiment tracking with MLflow. Model versioning with DVC. The full architecture and evaluation are documented in my university thesis.

Baby Monitor Pro 3.0

Started and created a next-generation baby monitor from scratch. Designed around an ESP32 with OV2640 camera on a custom PCB, all firmware written in C using Espressif’s IDF framework. Real-time audio and video streaming over WebRTC with a privacy-first architecture — no cloud relay, direct peer-to-peer connections between the monitor and the parent’s phone.

Built the companion Android application with BLE provisioning and live stream playback. Explored Wi-Fi HaLow (802.11ah) for extended range on sub-1GHz bands using Morse Micro MM8108 modules, including custom kernel module work on Android 14 with UNISOC T616 hardware.

Also built the cloud backend for audio/video IoT streaming using MQTT, RTMP, NGINX, and AWS, plus infant pose detection and baby cry classification models for smart alerting.

Location History & Route Matching

Rebuilt the location history feature for SoyMomo’s smartwatches, turning noisy GPS traces from children’s devices into clean, meaningful routes. Deployed a self-hosted Valhalla server for map-matching and implemented multiple clustering strategies to find the best route that matches the actual path the child took during the day.

Implemented and evaluated ST-DBSCAN (spatio-temporal density clustering), adaptive clustering that adjusts parameters based on GPS data density, stop-detection algorithms for identifying stationary periods, time-windowed clustering, and trajectory simplification. All with weighted centroid calculations based on GPS accuracy. Built a Next.js dashboard for visualizing and comparing experiments across algorithms.

HeyMomo — Voice & AI Assistant

Created a voice assistant for SoyMomo’s smart devices using GCP speech-to-text and text-to-speech, enhanced with word2vec, topic modeling, and clustering for better intent understanding. Later evolved into an LLM-powered conversational system backed by Supabase for persistent memory — the assistant remembers facts about each child (favorite colors, foods, school subjects) and can manage contacts, send messages, and initiate calls.

Also developed and published an Android audio recording library on Maven, standardizing device-side audio capture across SoyMomo products and reducing development time for audio features.

Tablet Platform & System Engineering

Built and maintained core Android functionality for SoyMomo’s tablet platform, including system-level services, screenshot and monitoring infrastructure, foreground services, overlay management, MediaProjection, and OEM/system-app deployment. Refactored the tablet application update system with Cloudflare, reducing data transfer costs and improving update efficiency across the production fleet.

Also shipped an image detection system using YOLO for identifying inappropriate visual content on the tablets.