Nolan Pozzobon

As an AI Fellow at Opteryx, I shipped a humanization feature into a production LLM pipeline that generates regulated pharmaceutical stability reports for paying enterprise customers.

I built a feature extraction system across five heterogeneous document corpora to model the stylistic and structural patterns of human-authored reports, which lifted reviewer acceptance by 40%. I also designed an evaluation pipeline that scores generated reports on factual correctness, logical consistency, register, and how well they reflect customer goals. This replaced a slow manual review process.

Quantified the limits of localized model editing using counterfactual arithmetic edits on Qwen2.5-1.5B. A sweep of 13 standard editing configurations showed that every method which successfully injects a false belief collapses neighbor arithmetic accuracy from 90% to under 20%.

We compared three preservation mechanisms (KL regularization over a 50-item multi-domain locality set, explicit neighbor replay, and a frozen-backbone correction head) across arithmetic, geography, and CounterFact edits. An inference-time “alpha sweep” over the LoRA task vector gives continuous post-hoc control of the edit-preservation tradeoff.

Engineered a multi-agent system using an AI architecture that integrates NLI nodes and RAG retrieval to mitigate “bullshit” assertions and promote nuanced social media speech.

The system gathers external context from sources like Wikipedia and BBC to perform structured reasoning. I validated performance across diverse agentic designs using models like Gemma 3, Llama 3.1, and GPT-OSS.

This research developed NLP techniques to better process and model informal language across languages.

We collected novel datasets for fine-tuning models using BeautifulSoup to more accurately model slang cross-linguistically. We also expanded and modified LLM architectures with PyTorch on distributed GPUs to build on previous research on embedding slang terms.

Research with Dr. Zhewei Sun at the Speech and Language group at TTIC.

Developed an AI native desktop application for real-time conversations with a personalized AI system to expand vocabulary using spaced repetition.

Crafted the workflows, prompts, and API endpoints, implementing a host of AI tools and APIs such as LangGraph, Whisper, and ElevenLabs.