Simon Aytes

About

I move fast from rough ideas to working AI systems, then do the engineering needed to make them reliable. That usually means working closely with stakeholders to understand what they actually need before touching code, and building something that can ship and stay there.

Experience

2021 – 2024, 2026 – Present

Penta Group

Senior Data Scientist / Data Scientist

Penta is a global strategy and communications consultancy where my work has moved from client-facing NLP and media-intelligence systems into company-wide AI tooling. Across both stints, the common thread has been embedded engineering: understanding analyst workflows, finding the right technical shape for ambiguous needs, and shipping tools that hold up in live client delivery.

• Built NLP and social-listening systems for media intelligence across social and traditional news data.
• Returned to design and ship internal AI tooling used by analysts across the company.
• Reworked notebook-era infrastructure into API-first services, worker execution, persistence, and observability.

Feb 2024 – Dec 2025

KAIST MLAI Lab

Graduate Research Assistant

At KAIST's MLAI Lab, I worked on efficient LLM reasoning and test-time compute. The work built out the research side of my practice — reading fast, testing ideas rigorously, and turning model-behavior questions into concrete methods evaluated against real reasoning workloads.

• First-authored Sketch-of-Thought, an EMNLP 2025 paper on efficient LLM reasoning.
• Developed an inference-time method that reduced reasoning tokens by roughly 84% on average.
• Explored training-free methods alongside GRPO and supervised fine-tuning approaches.

Jan 2025 – Present

Lighthouse Analytics

AI Engineering Consultant

Lighthouse Analytics is the name I use for independent AI engineering work. The main engagement was a production invoice-intelligence system for a small-business client, scoped from discovery with non-technical stakeholders through rollout. It used LLM-driven OCR, OpenAI structured outputs, and Pydantic schemas to turn messy vendor documents into dependable business data.

• Processed roughly 3,000 invoices per month across 100+ vendor formats.
• Saved an estimated 2,000 working hours and $30,000+ in annual operating costs.
• Helped the client scale a high-volume back-office workflow without adding headcount.

Aug 2021 – Dec 2021

The Wall Street Journal

Journo-Tech Fellow

At The Wall Street Journal, I worked on applied machine learning for editorial analytics. The fellowship was an early exposure to building against a real product environment: using internal content data, preparing training sets from production databases, and framing model outputs around decisions an editorial team might actually care about.

• Designed and trained an ML model to predict article performance based on its headline.
• Trained a model on 13,000+ metadata-rich articles from WSJ's content databases.
• Embedded with the News Insights team to develop data-driven solutions for journalists.

2019 – 2021

NASA Langley Research Center

Development Intern

Across several internships at NASA Langley, I built AR and VR tools for researchers working with 3D data. The work combined software development, data visualization, and user research: talking with researchers, understanding how they inspected technical data, and turning those workflows into interactive Unity prototypes.

• Led development of a VR heat-map visualization tool adopted by Langley's Engineering Design Studio.
• Interviewed 13 researchers to shape requirements and interaction design.
• Built Unity3D prototypes with integrated data processing, analysis, and visualization workflows.

May 2022 – Oct 2022

York College, CUNY

Researcher

At York College, I supported public-interest research on indigent burials by helping turn scattered public records into a structured research dataset. The project sat at the edge of data engineering and social research, with an emphasis on feasibility, source coverage, and making messy public information usable for analysis.

• Designed an automated scraping and extraction pipeline for state and local records.
• Helped assess the feasibility of a national database across jurisdictions.
• Contributed analysis supporting a Population Association of America research abstract.

Selected Work

Skills

Language Models

RAG·Agents & tool use·MCP·Prompt & context engineering·Structured outputs

Applied AI

Fine-tuning (SFT, GRPO)·Inference optimization·Anthropic & OpenAI SDKs

ML & Data

PyTorch·NLP·scikit-learn·pandas·NumPy·SQL

Engineering

Python·TypeScript·FastAPI·React·PostgreSQL·REST APIs·Git

Cloud & Infra

AWS (ECS, EC2, S3, SageMaker)·Docker·CI/CD (CodePipeline, CodeBuild)·Linux

Education

Contact

If you think we'd work well together, get in touch.