Tin Franovic

Building real-time voice AI and agentic systems powering automated call insights and compliance evaluation.

• Designed and implemented real-time low-latency voice call transcription that integrates with an event-driven backend to power automated call insights
• Developed a bidirectional communication pipeline with Amazon Bedrock AgentCore agents for real-time call compliance evaluation and analysis
• Helped bootstrap two engineering teams, setting technical standards and guidelines and developing AI tooling and processes for agentic development

Bridged Data Science and Engineering, delivering AI-powered features including RAG-based text summarization and narrative generation.

• Redesigned the AI extraction pipeline into an event-driven architecture, cutting end-to-end execution time by 75% (45 → 11 minutes) with zero-downtime migration
• Delivered a customer-facing AI narrative generation feature with a unified GenAI service supporting parallelism, caching, and sync/async invocation — adopted by all customers
• Built the interface between Data Science and Engineering for entity/customer data ingestion powering RAG features like summarization and outlier detection

Led engineering vision for Language-of-Preference within Amazon Search, scaling ML translation systems across 12 Retail stores worldwide.

• Integrated Search signals into ML model training, rebuilding 14 language models and measurably improving customer experience
• Launched 33 languages of preference across 12 Amazon Retail stores through cross-team collaboration spanning language detection, spell correction, and machine translation
• Drove adoption of a new ML training framework, achieving 85% reduction in model training costs through modularization and AWS Batch/ECS optimization
• Built a highly scalable translation cache for frequently searched terms, with self-refreshing and emergency override reducing correction time from days to 15 minutes
• Managed $1.3M annual hardware budget, driving efficiency programs that absorbed rising GPU demand without cost overruns
• Mentored engineers toward senior promotions and led org-wide initiatives on engineering excellence and deployment best practices

Designed and scaled Amazon's core machine translation API from inception to a Tier-1 service processing 500M+ requests weekly.

• Led design of the internal machine translation API, growing it to hundreds of clients and evolving it into a Tier-1 service — still Amazon's primary translation interface
• Scaled the API for 100x traffic increase to support Language of Preference on Amazon Retail, implementing A/B testing, caching, client isolation, and throttling
• Managed ecosystem integration for 2 acquisitions, 3 external MT providers, and a third-party code scanning tool, collaborating with Legal on due diligence

Research engineer in computational neuroscience, optimizing neural network simulations for supercomputer architectures.

• Published "Cortex-inspired network architecture for large-scale temporal information processing" at Neuroinformatics 2013
• Implemented high-performance C/C++ neural network algorithms optimized for Cray supercomputers
• Overhauled floating-point arithmetic in a neural simulator, achieving 10x throughput increase
• Introduced GitLab, CI/CD, and testing practices to the lab — later adopted by other university labs