SENIOR SOFTWARE ENGINEER

Nathan
El-Khoury.

I build AI agents, automation pipelines, and core infrastructure at Byzfunder.

View experience Résumé
Currently at Byzfunder
Senior Software Engineer - AI Agents & Automations
PERSONAL INFORMATION PROCESSORN.E. MODEL 01
INITIALIZING DEVICE…
NATHAN-OS ONLINE
PERSONNEL FILE01
whoami

Nathan El-Khoury

Senior Software Engineer

AI Agents & Automations

Ottawa, Canada

 

Currently → Byzfunder

Type help to look around. Make yourself at home.

Explore a section. Try the dials.Dials: skills / contact
N.E. MODEL 01 / PERSONAL INFORMATION PROCESSOR

↑ ↓ command history Tab autocomplete Esc close

↳ OTTAWA, CANADA
GitHub LinkedIn SCROLL TO EXPLORE

Experience

August 2026 — Present CURRENT

Byzfunder

Senior Software Engineer - AI Agents & Automations

Building core infrastructure, portals, and automation pipelines for Byzfunder’s centralized lending platform, designed to support millions of dollars in financing.

Sept 2025 — August 2026

J-Squared Technologies

AI Engineer · FALC AI Team

  • Built and shipped an internal agentic platform used across teams to run complex multi-agent workflows, with typed agent contracts, per-agent MCP scope enforcement, budget + access guardrails, and a two-tier kill switch
  • Eliminated data loss on mid-run failures by designing durable workflow orchestration with crash-safe checkpointing and human-in-the-loop interrupts, so runs resume exactly where they left off after hub restarts
  • Cut candidate screening from days to minutes and saved recruiters hours per role with a production AI system that has ranked 2000+ resumes via a novel two-track engine combining LLM keyword confidence scoring with qualitative evaluation
PythonLangGraphFastAPINext.jsPostgres
May 2025 — August 2025

Trend AI

Software Engineer Intern · Reports Team

  • Led development of a multi-agent LLM compliance system converting natural language queries into on-demand visual reports for enterprise customers, cutting delivery time from weeks to under 30s
  • Autonomously resolved 90% of user queries by orchestrating a LangGraph workflow spanning NLP, MCP-based data retrieval across 4 servers and 60 APIs, and LLM-driven visualization with auto-generated PDF output
  • Ingested raw JSON into an in-memory relational database with Pandas and SQL, enabling an LLM-SQL agent to fetch visualization data in under 4s and reduce context windows by 80%
PythonFastAPILangGraphMCP
May 2024 — May 2025

J-Squared Technologies

Machine Learning Engineer Intern · FALC AI Team

  • Secured $100k+ in contracts through live conference demos of a CNN-LSTM rockfall classification system achieving 92% accuracy under varying environmental conditions, built end-to-end from ETL pipeline to deployed model
  • Boosted model accuracy 25% and cut false positives 30% by building a Python synthetic data generation framework that expanded training data variety for the computer vision team
PythonPyTorchTypeScriptNext.jsNestJS
Sept 2023 — Dec 2023

Apption Software

Software Engineer Intern · Datahunter AI Team

  • Architected the Datahunter Query Language using a command-pattern backend integrated into the Blazor frontend, accelerating user query resolution time by 200%
  • Boosted enterprise document processing throughput 5x by implementing NLP-driven keyword extraction with TF-IDF and word co-occurrence analysis
C#PythonBlazor
May 2023 — Aug 2023

Starvoy Technologies

Software Engineer Intern · Hardware Team

  • Achieved 98% accuracy and cut manual QA time 30% by developing an SVM-based anomaly detection system with k-fold cross-validation for optical transceiver testing
  • Saved 15 engineering hours/week by automating optical transceiver report generation with LLM pipelines, replacing a manual reporting workflow
PythonC++TypeScript

Projects

Retail Trading Journaling Platform

Full-stack trading journal built for data-driven retail traders — customizable dashboards, granular trade tagging, and analytics aimed at helping traders refine their edge. Extended with agentic daily market debriefs and a personalized AI coach that analyzes user trade history to surface behavioral patterns and actionable feedback. Currently serving a small private user base.

Next.jsTypeScriptFastAPIPostgreSQLLangGraph

Librarian — Personal Agentic Knowledge System

Obsidian-style web client paired with a Librarian agent that reads, organizes, and surfaces notes on demand, with Speechify for text-to-speech playback.

LangGraphNext.jsTypeScriptSpeechify

Machine Learning UFC Predictor

Source

Engineered features and trained an ensemble ML model (XGBoost + Logistic Regression) on historical fight data, deployed via ReactJS + Flask for real-time predictions at 86% accuracy

PandasScikit-learnSQLFlaskReactJS

Computer Vision Poker Game Theory Optimizer

Source

Built a YOLOv8 + ByteTrack card tracking pipeline streaming detections into Monte Carlo simulations to compute win odds and recommend optimal moves (fold / call / raise)

PythonUltralyticsSupervisionOpenCV

MRI Brain Tumor Detection

Source

Developed a CNN-based tumor detection system matching DenseNet’s 86% accuracy at 1/3 the computational cost.

PythonUltralyticsOptunaScikit-learn

Technical skills

AI SYSTEMS
INTELLIGENCE Data Models Agents Production

Languages

PythonTypeScriptC++C#JavaSQLBash

AI / ML

PyTorchTensorFlowScikit-learnLangGraphLangChainOpenAI APIMCPPandasNumPyPydantic

Frameworks

FastAPIFlaskNestJSNext.jsNode.js.NETReact

Databases

PostgreSQLMongoDBPineconeChromaDBNeo4jFAISS

Cloud / DevOps

DockerKubernetesAWSLinuxGitGitLab CI

Research

THE EXPERIMENT
Single passMultistage fusion ↗Multi-agent debateHybrid retrieval
0.800F1 · MULTISTAGE FUSION73%TOON TOKEN SAVINGS*

*At a 0.067 F1 cost relative to JSON.

Course Project / 2026LLM SYSTEMS

Empirical Evaluation of LLM Pipeline Architectures and Output Formats for Automated Candidate Screening

Nathan El-Khoury · Carleton University — COMP 4900 (GenAI)

Controlled comparison of four LLM pipeline architectures (single-pass, multistage fusion, multi-agent debate, hybrid retrieval) for ranking 100 production resumes with GPT-4.1. Multistage fusion wins (F1 = 0.800, Spearman ρ = 0.896); multi-agent debate underperforms the single-pass baseline. Also first empirical evaluation of TOON vs. JSON for structured output — 73% token savings at a 0.067 F1 cost with 100% parse reliability.

Read the paper

Education

2022—2026OTTAWA, CANADA

Carleton University

Honours BCS. of Computer Science · AI & ML Stream

Artificial IntelligenceMachine LearningNeural NetworksDatabase ManagementData Structures and AlgorithmsAlgorithm Analysis and DesignOperating SystemsStatisticsLinear Algebra
3.97/ 4.0 GPA