Erina Karati

Erina Karati is an applied AI engineer and co-founder of MinneDigest, a platform that turns local reporting into multilingual news summaries and audio. Her work spans publishing tools, autonomous game characters and private health assistants, with a recurring concern: how AI systems remember information, protect it and use it to make decisions.

From enterprise networking to applied AI

Karati spent three years at Microsoft, from September 2021 to August 2024, debugging enterprise Windows networking systems. She investigated failures across networking, identity and security infrastructure through packet captures, protocol traces and root-cause analysis, and developed PowerShell automation for recurring problems. That work shaped her approach to observability and reliability in AI systems.

She then pursued a master’s degree in computer science at the University of Minnesota, graduating in May 2026. Her graduate work included research and teaching in machine learning, natural-language processing and generative AI. In 2025, she built multi-agent systems at Supercell’s AI Innovation Lab and subsequently worked at LyftBridge Innovation, integrating real-time conversational agents with cloud APIs and Google Workspace workflows. Her engineering background and project portfolio trace this move from diagnosing distributed failures to building systems whose behavior depends on model-generated decisions.

Making local reporting easier to use

Karati created MinneDigest with Arunachalam Manikandan and Dipan Bag while the three were graduate students. Their prototype won the Minnesota Journalism Center’s AI and journalism hackathon in early 2025, earning a $10,000 award for further development.

The team combined summaries of Minnesota reporting with translation and podcast-style audio, giving readers different ways to follow civic information. In an early-2025 interview about their product vision, the team described personalized recommendations and direct collaboration with news outlets as next steps.

Memory that changes behavior

At Supercell’s AI Innovation Lab, Karati and Manikandan developed Project Paradox, a framework for autonomous game characters that interact, compete and cooperate with players and one another. Their initial exploration of a social-deduction game grew into a more general system for interactive characters.

Each agent has a separate retrieval-backed memory, emotion values and trust scores. Events and conversations update these states, while importance scoring helps retain significant memories for later retrieval. Agents can plan actions, manipulate objects and respond to encounters according to their own experience. Because they lack an omniscient shared memory, information must travel through conversations.

Karati’s question is whether memory, emotion and trust become active decision signals: do they change what a character does, whom it believes and how it responds to later events? The collaborators’ abstract on these mechanisms in social agents was accepted to the WiML Symposium at ICML 2026.

Her work develops three related ideas:

• Source-aware agent memory: Remembering a topic does not ensure that an agent remembers its origin or uncertainty. In Project Paradox, a rumor about a mango sale lost context as agents passed it along over extended interactions. Karati’s approach to long-horizon agents calls for preserving whether information is firsthand, secondhand, verified or uncertain, and distinguishing remembered events from current beliefs.

• Behavioral evaluation across whole runs: Karati proposes controlled scenarios and structured traces to test information reach, source retention, uncertainty, replanning and privacy together. An experimental loop can adjust selected memory, communication, trust and planning policies while keeping the evaluation harness, scenarios and metrics fixed. Changes survive only when the scorecard improves and safeguards hold; otherwise, they are reverted. This matters because faster information sharing can also cause oversharing, while better recall can increase reliance on stale memories. She reports a more contextual response in the mango example, but withholds claims of general improvement pending repeated results.

• Privacy through architecture: For Project Sirius, an on-device health assistant built with Bag and Manikandan, Karati owned the Health Memory Module. Her account of its design describes local storage, structured and semantic retrieval, limited context assembly, encrypted device synchronization and prompt-injection defenses. Medication and allergy questions use structured retrieval rather than relying entirely on similarity search. Keeping sensitive context on the device shapes how the system stores, retrieves and shares information from the beginning.