Erina
Karati

AI researcher · ML engineer · Graduate student

M.S. Computer Science student at the University of Minnesota, graduating May 2026, focused on agentic AI and generative AI systems. Previously built multi-agent frameworks at Supercell, integrated real-time conversational AI at LyftBridge, and spent three years at Microsoft debugging Windows networking at enterprise scale.

Resume Read Blog

Technical taxonomy

erina@skills:~$ tree ./taxonomy
▾ ai-ml/
agentic-ai multi-agent-systems generative-ai llms rag-systems deep-learning nlp mcp
▾ languages-libraries/
python java c / c++ / c# powershell r tensorflow pytorch scikit-learn hugging-face langchain langgraph fastapi
▾ cloud-devtools/
azure azure-ml gcp aws active-directory linux git / github postgresql / sql vmware / hyper-v wireshark
▾ systems-networking/
distributed-systems rest-apis client-server-architecture wireless-networking tcp-ip computer-networks-security

Work experience

Industry experience

  1. Jul 2025 – Aug 2025

    GenAI Intern · LyftBridge Innovation

    Minneapolis, MN

    • Built an agentic AI assistant with modular architecture, streamlining tool invocation flows and improving system reliability.
    • Integrated ElevenLabs real-time conversational agents with Google Cloud Run APIs (FastAPI + Zapier MCP) to automate Google Workspace workflows, reducing response latency by 70% over prior local deployment.
  2. Apr 2025 – Jun 2025

    AI Engineer · Supercell

    San Francisco, CA

    • Developed a modular multi-agent framework using LLaMA 3, Agentic RAG, FastAPI, and MCP, achieving <100ms latency. Enabled dynamic agent planning, collaboration, personality, and emotion systems on local/cloud endpoints.
    • Orchestrated a scalable per-agent FAISS-based vector memory system supporting 1M+ semantic embeddings, analyzing retrieval quality, memory consistency, and latency–recall–cost tradeoffs with cross-agent information grounding.
  3. Sep 2021 – Aug 2024

    Windows Networking Engineer · Microsoft

    Bangalore, India · 3 years

    • Diagnosed production failures for 200+ enterprise clients (Tesla, Bank of America, UK Government, Stanford Medicine, PepsiCo), ensuring reliability and security across large-scale distributed systems.
    • Performed root-cause analysis across TCP/IP, Wireless, VPN, SDN, Load Balancing, Firewall, Active Directory, Azure — using network trace analysis, packet captures, and protocol-level debugging. Drove architecture improvements through PowerShell automation across multi-vendor infrastructure (Cisco, Palo Alto, VMware, Juniper, Aruba).
    • Subject Matter Expert in Wi-Fi. Recognized with the Global CSS Impact Award ($1,000) and Star ACE Team (Top 5%) for FY22 & FY23.
  4. Apr 2021 – Jul 2021

    Windows Engineer Intern · Microsoft (Aspire University Hire)

    Bangalore, India

    • Trained on Windows Setup, Performance, and Networking; Azure Cloud; Active Directory; and Microsoft business & culture values.

Academic experience

  1. Jan 2025 – Present

    Graduate Research & Teaching Assistant · University of Minnesota

    Minneapolis, MN

    • Developing GenAI automations and web-based tools for educators and researchers; built NLP pipelines over 500K+ research paper abstracts using RAG + LangChain + Word2vec, pushing trend-detection accuracy to 85%.
    • TA for IDSC 4504 — Machine Learning & Responsible AI for Business (120+ students); contributed to MSBA 6155 — GenAI for Business with tutorials on Azure AI Foundry, Copilot Studio, AutoGen, and multimodal RAG pipelines.

Education

  1. Aug 2024 – May 2026

    M.S. Computer Science

    University of Minnesota, Twin Cities · Minneapolis, MN

    GPA 4.0/4.0. Coursework: AI, Advanced ML, NLP, LLMs, GenAI.

  2. Aug 2017 – May 2021

    B.Tech, Computer Science and Engineering

    Techno India University · West Bengal, India

    GPA 9.11/10.

Projects.

Engineering at the intersection of agentic AI, generative systems, and the infrastructure that makes them reliable. Current focus: multi-agent orchestration, RAG at scale, and cost-efficient reasoning with smaller open-weight models.

MinneDigest

AI-powered news-to-audio platform for personalized multilingual digests

MinneDigest is an AI-powered news-to-audio platform built in under 8 hours during a university hackathon. It delivers real-time, personalized news digests through multilingual text and podcast-style audio. The project secured a $10,000 grant from Hacks/Hackers and the University of Minnesota to add features, scale the platform, and partner with local outlets such as MPR and MinnPost.

The system includes an end-to-end pipeline leveraging LLMs for summarization, real-time translation, and sentiment-aware text-to-speech, improving accessibility and engagement. A dual-persona podcast format lets AI agents discuss daily news conversationally, while the web interface lets users browse articles, select languages, and listen on any device.

  • FastAPI
  • Ollama
  • LLaMA 3
  • ElevenLabs
  • Whisper
Project Site → MinnPost Coverage →

Project Paradox

Multi-agent LLM framework for emotionally driven gameplay agents

Project Paradox is a multi-agent LLM framework developed as part of the Supercell AI Innovation Lab. It creates intelligent, emotionally driven agents for immersive gameplay, with each agent equipped with memory, beliefs, personality, and autonomous planning for lifelike interactions and emergent storytelling across any game world.

The framework combines a RAG-based memory engine, emotion-belief-personality modeling, spatially aware world understanding, contextual conversation, intent recognition, and a dynamic story trigger system for global event propagation. A universal game prompt layer lets designers shape distinct gameplay styles with simple prompts while keeping the system portable across engines, devices, and deployment environments.

  • Multi-Agent Systems
  • LLaMA 3
  • Agentic RAG
  • FastAPI
  • MCP
  • FAISS
Demo at Supercell HQ, Helsinki →

Agentic Code Repair

Multi-agent feedback routing for iterative debugging

A multi-agent code repair system built on Qwen2.5 that routes feedback between specialized agents (planner, executor, critic) to iterate on buggy programs. Each agent is narrow in scope but collectively they produce reasoning traces that match much larger single-pass models.

On a benchmark of 400+ LeetCode problems, a 7B-parameter architecture running across agents matched the accuracy of a single 32B model (78% solve rate) while reducing inference cost by 75%. Targeted failure-mode analysis drove the optimization.

  • Python
  • Qwen2.5
  • Hugging Face
  • Multi-Agent Systems
  • LLM Evaluation

NLP Pipelines for Scientometrics

Graduate Research Assistant · University of Minnesota

Large-scale NLP pipelines evaluating over 500,000+ research paper abstracts to quantify scientific growth and detect emerging research trends. The system uses RAG over the abstract corpus with LangChain plus Word2vec embeddings for semantic clustering.

Error analysis is the heart of the work. Most false-positive trends were explained by small-corpus effects in adjacent domains. Addressing these pushed trend-detection accuracy to 85%.

  • Python
  • RAG
  • LangChain
  • Word2vec
  • NLP

DNS Tunneling Detection

Classical ML comparison across real-world attack datasets

A detection system for DNS tunneling cyber-attacks that benchmarks Neural Networks, SVM, Logistic Regression, Random Forest, and K-Means across real-world datasets. The study maps where each method over- or under-performs, including boundary cases that matter for deploying classical ML in a security pipeline.

  • Python
  • TensorFlow
  • scikit-learn
  • Cybersecurity

Get in touch.

Always happy to talk about agentic AI, RAG systems, grad-school research, or building things at the intersection of ML and production infrastructure. The fastest way to reach me is email.