James Franzo

• • Architected and scaled the real-time AI inference pipeline serving 3B+ daily multimodal interactions (text, voice, image, video) across 200+ countries, contributing to MAU growth from 11M to 29M+ during tenure

• • Led full-stack development of the Next.js web platform and companion features, including voice chat with ultra-low latency in 40+ languages, character creation tools, and a gamified collectible card system that drove a 2.5× increase in average revenue per paying user ($6 -> $15 ARPPU)

• • Integrated successive generations of proprietary LLMs (ABAB 6.5 through M2.5, 230B parameter MoE architecture) into production, coordinating model rollouts for Talkie's conversational engine while maintaining sub-200ms response latency at scale

• • Built multimodal content generation features, AI image generation (20M+ images/day platform-wide), text-to-speech integration (Speech-02 model, 70K+ hours synthesized daily), and video generation, enabling Talkie to differentiate as the only companion AI platform with full text/voice/image/video/music capabilities

• • Developed a proprietary recommendation engine combining LLM-driven natural language understanding with collaborative filtering and real-time pricing APIs (Skyscanner, Booking.com, GetYourGuide), enabling personalized multi-destination itinerary generation with live availability and fare data

• • Built computer vision pipelines to ingest, label, and index 350,000+ destination videos from 1,600+ creators, automatically extracting up to 20,000 data points per video to power visual search features (e.g., "show me destinations that look like Mars") and content-to-booking pairing

• • Engineered the Instagram DM chatbot using Meta's Messaging API, extending the AI travel agent to the Beautiful Destinations audience of 50M+ social followers and creating a novel distribution channel that drove early user acquisition at near-zero cost

• • Designed the flight price prediction model and PriceLock algorithm, training time-series forecasting models on historical fare data to advise users on optimal booking windows, a key differentiator highlighted in seed-round pitch materials that helped close a €3M raise from Firstminute Capital, M13, and strategic angels

• • Implemented the multi-language NLP stack supporting 12 languages at launch, including intent parsing, entity extraction for locations/dates/budgets, and a custom-tuned conversational persona, reducing average trip planning time from 10+ hours to under 15 minutes

• • Developed semantic segmentation models (U-Net, DeepLab architectures) for automated sky detection and replacement, BoxBrownie's most popular enhancement feature achieving 95%+ accuracy on complex architectural edges and reducing manual editing time by 40% across millions of processed images

• • Built deep learning pipelines for day-to-dusk photo conversion using GAN-based image-to image translation, automating realistic twilight lighting simulation that previously required 15+ minutes of manual artist work per image, enabling same-day turnaround at scale

• • Engineered computer vision systems for room classification, object detection, and spatial analysis that powered automated routing of incoming jobs to specialized editors, virtual furniture scaling for staging accuracy, and quality-control validation across 50,000+ customer accounts in 82+ countries

• • Designed and deployed the image processing backend supporting the company's REST API (launched Q12019), enabling enterprise clients like bridgeMLS, the first U.S. MLS integration, to programmatically submit and retrieve edited photos at 100 requests per 10 seconds