Erina Karati builds applied AI systems, including multi-agent game companions and tools for local journalism. Her work on Project Paradox examines how agents preserve memories, uncertainty and the consequences of earlier interactions.[1][2][3]
Start with Long-Horizon Agents Need Experiments, Not Just Prompts — Erina Karati →
GitHub Background
Karati earned an M.S. in Computer Science at the University of Minnesota in May 2026 after a B.Tech in Computer Science at Techno India University. Her CV describes software engineering at Microsoft from September 2021 to August 2024, diagnosing Windows and networking failures. Later work included multi-agent systems at Supercell and a conversational assistant at LyftBridge Innovation.[4]
Contributions
News that can be read or heard
With Dipan Bag and Arunachalam Manikandan, Karati built MinneDigest, a local-news digest and podcast prototype. The University of Minnesota describes retrieval from Minnesota news outlets, summaries and multilingual audio. The team won its AI and journalism hackathon and a grant to continue development.[3]
Memory needs source and certainty
Project Paradox gives game companions separate memories, emotions, beliefs and plans. In her talk, Karati shows how a rumor can retain its topic while losing who said it or whether it was certain. Her experimental approach evaluates whole runs using controlled scenarios and structured traces of conversations, memory operations and actions. A revised policy must preserve attribution and improve behavior without violating privacy or other constraints.[2]
Ideas
Karati separates a character’s local knowledge from an evaluator that inspects the scenario. It checks whether agents share public facts, preserve uncertainty and revise plans. Freezing evaluation while changing a small policy area permits rollback. One successful run does not establish general improvement.[2]
Tools, writing & talks
Writing, research & talks
Start here
A game-agent demonstration and experimental design for testing memory, attribution and planning across a run.
Start with the rumor example to see why retrieving a memory is insufficient if its source and certainty disappear. Some familiarity with agents helps; the later scorecard gives a practical way to design your own long-run tests.
Speaker[2]Original source- MinneDigest project accountcse.umn.edu
The university explains the team-built local-news prototype and its hackathon origins.
Co-builder profiled by university[3]