Patrick Serrano

Patrick Serrano

Applied AI engineer & technical product lead

I build retrieval-augmented generation systems end to end: corpus ingestion, retrieval design, evaluation, and the product around them. I have done it both as the engineer writing the system and as the person owning product direction and users.

Los Angeles, CA Email me LinkedIn

Experience

SerraiLabs

Founder / Head of Engineering

– present · Los Angeles

A legal research platform for the Supreme Court of the Philippines, a court system that had almost no research infrastructure of its own. Discovery with the court’s CTO, Deputy Clerk of Court, and working attorneys isolated the core problem: researchers had no reliable way to tell whether a cited precedent was still good law.

  • 66,000court decisions in the citation-treatment corpus
  • ~19%systematic mis-citation error eliminated
  • 50+court researchers and attorneys using it in production
  • Built a citation-treatment engine that classifies how each case cites another (affirmed, distinguished, overruled) and surfaces forward-citation maps, backfilling thousands of corrected authority links.
  • Architected the retrieval stack: hybrid search plus knowledge graphs mapping legal relationships across decades of precedent, over a proprietary ingestion and indexing pipeline.
  • Deployed on serverless AWS (CloudFront, Lambda, S3, Bedrock) with Cognito SSO, per-user data isolation, and private document ingestion, meeting a judicial institution’s confidentiality requirements at near-zero idle cost.
  • Owned product direction from MVP to launch, including three rounds of pilot testing with court research staff that shaped the UI and the roadmap.

SDSU Research Foundation / Carnegie Mellon

Research Lead

–

A program spanning COVID-19 epidemiological modeling, AI for the U.S. Navy, applied LLM work, and DOE Project Genesis.

  • Led a team of 12 researchers, coordinating with scientists at Lawrence Livermore and Argonne National Laboratories, and fine-tuned LLaMA-3-70B for literature-review parsing and document clustering.
  • Developed a two-step LLM approach to document clustering that matched or beat traditional methods in every test, with superior metrics in over half of evaluations. Patent pending with the Department of Energy.
  • Designed the architecture and nine-month plan for a QLoRA-tuned LLM refiner with panel-disagreement calibration that abstains on low-confidence verdicts, aimed at DOE exascale codebases.

Tata Consultancy Services

Capstone: LLMOps / MLOps for GenAI

– · Pittsburgh

  • Architected an enterprise-grade LLMOps platform with automated deployment pipelines and secure API integration.
  • Prioritized observability, on the finding that teams will not trust models they cannot monitor: real-time performance tracking, drift detection, and resource dashboards in a Streamlit interface that non-engineers could read.
  • Fine-tuned and deployed LLaMA 3.1 8B models for enterprise log analysis, enabling RAG, anomaly detection, and root-cause analysis over large-scale system logs.

O.E.M Financial and Investment Services

Financial Analyst, later Business Analyst Intern

– · Diamond Bar, CA

  • Built and presented Los Angeles real estate market analyses to firm leadership that informed future investment decisions. Returned in summer 2024 for the internship required by the Carnegie Mellon program, assessing potential returns on L.A. real estate investments.

Publications & recognition

Publications

  1. LLM Driven Document Clustering: Improving Real-time Security Intelligence and Threat Detection

    First author. IEEE International Conference on Intelligence and Security Informatics (ISI), Hong Kong.

  2. Curve4Covid: Comprehensive Understanding via Representative Variable Exploration

    First author. Americas Conference on Information Systems (AMCIS 2021). Top 25% of accepted papers.

Patent & awards

  • Pending

    Patent, U.S. Department of Energy

    LLM-based document clustering, from the research program above.

  • Patient Safety Fellowship, Jewish Healthcare Foundation

    A computer vision framework for STAT MedEvac that transcribes real-time events during helicopter emergency missions.

  • Finalist, Hack-a-Startup

    An LLM web app that simplifies complex text for readability.

  • Winner, The Pitt Challenge

    Healthcare hackathon. An application targeting user-specific tinnitus frequencies.

Skills

Applied AI
RAG, hybrid search, knowledge graphs, evaluation design, LLM fine-tuning (QLoRA), LLMOps / MLOps, drift detection, document clustering, anomaly detection, computer vision, modeling and forecasting
Languages & libraries
Python, SQL, Java, Bash. PyTorch, TensorFlow, Keras, scikit-learn, LangChain, NLTK, Gensim, NumPy, Pandas, Flask, Streamlit
Cloud & data
AWS (Bedrock, Lambda, S3, CloudFront, Cognito), GCP, Docker, Linux, REST APIs, Ollama. PostgreSQL, MongoDB, Tableau, SAP ERP
Product
Product discovery, requirements scoping, roadmap ownership, pilot and user testing, technical presentations to leadership, management consulting

Education

  • Carnegie Mellon University

    MS, Information Systems Management, Business Intelligence & Data Analytics

    – · Pittsburgh

  • San Diego State University

    BS, Management Information Systems, minor in Computer Science

    – · San Diego

Languages: English, Filipino / Tagalog, and basic Spanish.

Contact

pat@pserrano.net