Alan Mohan

Alan Mohan

AI engineer, pursuing an M.S. in Artificial Intelligence Engineering – Information Security at Carnegie Mellon University, expected December 2026.

I work on what happens after the demo: whether an AI system does what it claims, why it produces what it produces, and how it holds up when someone attacks it. In practice that means agent evaluation, interpretability and generative-AI security.

View projects alanmohan@cmu.edu

Alan Mohan, photographed from the shoulders up in a white button-down shirt against a plain light-grey backdrop.
Pittsburgh, 2025.

Skills

Programming
Python (primary), Java, JavaScript, PHP, C/C++, SQL
AI/ML and generative AI
Generative AI, large language models (LLM), prompt engineering, explainable AI (XAI), PyTorch, Keras, LangChain, vector databases, retrieval-augmented generation (RAG), HuggingFace, MCP, LLM interpretability, LLM evals
Cloud infrastructure
AWS, GCP, Docker, CI/CD, Git, Linux, Flask

Projects

Experience

  1. May 2026 – August 2026Pittsburgh, PA

    AI InternBank of New York

    I built the agent-evaluation platform for the firm's AI Hub, so that judging an agent stopped being a manual, one-off exercise. The evaluation engine carries more than thirty metrics — RAGAS, LLM-as-a-judge, and multi-turn agent measures for goal completion, knowledge retention and tool-use accuracy — behind a single workflow covering dataset management, synthetic test-case generation, online evaluation and performance analysis.

  2. January 2026 – May 2026Pittsburgh, PA

    Graduate Research AssistantCarnegie Mellon University

    I wanted to know where inside a language model a financial comparison actually gets decided. I extracted activations across all 28 layers over more than five thousand mutual-fund comparison samples and trained linear probes on them. Input features were recoverable above 90% accuracy, with encoding peaking mid-stack, but the probe for the model's final decision trailed the probe for its inputs by 6.3 points — a measurable gap between what the model has represented and what it does with it.

  3. February 2025 – July 2025Dubai, UAE

    AI InternCondo Protego

    Sales proposals were assembled by hand from scattered price lists and specifications. I built a Flask and React application with generative-AI integration, deployed on Google Cloud Run, that cut proposal creation time by 70%, plus a set of automation agents for data extraction and PDF processing that returned about twelve hours a week to each person using them.

  4. June 2023 – August 2023Remote

    Generative AI and Backend Developer InternDigiProctor

    Medical referrals arrived as scanned documents, often handwritten, and were triaged by a person reading each one. I built a prioritisation system in Python that used AWS Textract to pull structure out of the scans and GPT-3.5 to judge urgency, taking a document from roughly ten minutes to about thirty seconds. I also worked on the web application in PHP CodeIgniter and jQuery.

Contact