Resume - Andreas Atle

Houston, TX (Remote)

• Phone: (832) 488‑9813
• Email: atle.andreas@gmail.com
• GitHub: https://github.com/andreasatle
• LinkedIn: https://linkedin.com/in/andreasatle/
• Website: https://atle.dev

Senior AI Engineer — Agentic Systems

Senior AI Engineer with extensive experience building production‑grade, governed AI systems. Specialized in LLM‑based agent architectures with strict schema validation, bounded execution, and enterprise‑ready safety controls. Proven ability to design, ship, and iterate on AI agents that integrate into real business workflows with measurable impact, cost awareness, and security.

Agent Architecture Focus: planner–worker–critic loops, supervisor orchestration, schema‑enforced I/O, bounded retries, deterministic state transitions, and auditable outputs.

Core Skills

AI & Agents LLM‑based agents (OpenAI SDK), RAG architectures, embeddings, evaluation harnesses, agent governance & safety

Programming Python (primary), Go, C/C++, SQL

Frameworks & APIs FastAPI, Flask, gRPC, REST, Pydantic v2

Infrastructure & Data AWS, Docker, Kubernetes, ChromaDB, MySQL, SQLite

Practices Schema‑driven design, cost‑aware inference, deterministic workflows, CI/CD, Git

Professional Experience

Independent AI Engineer — Enterprise Agents

Houston, TX · 2021–Present

• Designed and deployed production LLM agent pipelines using OpenAI Agent SDK and strict Pydantic schemas, processing 1,300+ legal documents with deterministic validation, bounded retries, and auditable outputs, reducing manual review effort by 70%.
• Implemented agent governance safeguards, including planner isolation, schema enforcement, and explicit ACCEPT/REJECT evaluation steps to prevent hallucinations, prompt injection, and uncontrolled state mutation.
• Built embedding‑based RAG systems (ChromaDB) enabling secure, structured retrieval across legal and financial corpora, optimized for accuracy and inference cost.
• Collaborated with non‑technical stakeholders to redesign document‑review workflows with agents at the center, balancing automation gains with operational reliability.

Software Developer — Spectacle LLC

Remote · 2022–2023

• Deployed ML models (BERT‑based classifiers, translation, content moderation) into AWS EKS, supporting production workloads.
• Built high‑throughput APIs using gRPC and Python, emphasizing reliability, observability, and operational correctness.

Scientific Researcher — Total E&P USA

Houston, TX · 2009–2020

• Developed high‑performance numerical algorithms for seismic imaging and inversion, deployed on HPC clusters.
• Accelerated production workflows using MPI and CUDA, emphasizing numerical stability, performance guarantees, and reproducibility.

Postdoctoral Researcher — Memorial University of Newfoundland

St. John’s, Canada · 2006–2009

• Developed advanced finite‑difference and eikonal solvers adopted in seismic inversion workflows.

Selected Projects

Enterprise Agentic Workflow for Legal Documents OCR → schema‑validated JSON → RAG → structured outputs pipeline using OpenAI SDK and Pydantic, designed for auditability and cost control.

Education

• Ph.D., Applied Mathematics — Stockholm University
• M.Sc., Scientific Computing — KTH Royal Institute of Technology
• B.Sc., Computer Science — Stockholm University

Certifications

• AWS Certified Cloud Practitioner
• Professional Scrum Master I (PSM I)

Profile - Andreas Atle

I work on controlled, production-grade AI systems, with a focus on agentic architectures that are constrained, auditable, and predictable.

Lately, my work has centered on designing and implementing a schema-driven document system built around explicit intent envelopes, deterministic agent execution, and strict validation boundaries. The goal is not autonomous behavior, but reliable transformation of structured human intent into structured outputs, suitable for real workflows.

Recent work includes:

• Building an intent-driven document writer with explicit Planner–Worker–Critic separation and schema-validated I/O.
• Designing UI and backend contracts that favor clarity over abstraction, avoiding hidden framework behavior and layout drift.
• Iterating on agent governance patterns: bounded retries, explicit failure modes, and zero silent degradation.
• Starting a PDF → legal-text extraction pipeline, focused on best-possible fidelity, explicit uncertainty tracking, and disciplined OCR correction rather than “helpful” guessing.

Across projects, the emphasis is consistent:

• Determinism over cleverness
• Explicit contracts over inference
• Systems that are understandable months later, not just impressive once

I am interested in work where engineering discipline matters, especially at the boundary between human intent, automation, and accountability.