Reza Arabsheibani
Melbourne, Australia

Reza Arabsheibani, PhD

Applied AI Engineer·Machine Learning·PhD in Deep Reinforcement Learning

I build AI systems that hold up in production. LLM pipelines where every model output is treated as untrusted until it passes automated evaluation gates and human review, reinforcement learning agents trained on GPU clusters, and ML models that delivered measurable savings. My PhD at the University of Melbourne trained deep RL agents to follow natural-language instructions; today I bring that rigour to production AI on Azure.

See my work LinkedIn GitHub Email
10+ yearsdata & ML engineering
PhDDeep RL & spatial AI
6peer-reviewed papers
775LLM-translated objects shipped

Selected work

Production AI, research and applied ML. Each project solved a real constraint: risk, reliability, scale or messy data.

Applied AI · Parks Victoria

Evaluation-gated LLM code migration

An AI-assisted, test-driven pipeline that translates legacy Oracle PL/SQL and Oracle Spatial logic to PostgreSQL/PostGIS. Model output is never trusted by default.

  • Microsoft Foundry-hosted models and GitHub Copilot CLI, with retrieval over a curated migration knowledge base and strict JSON response contracts
  • Every candidate must pass syntax, dependency, structural, spatial, behavioural and security gates, then human review
  • Bounded retries, immutable evidence, resumable state; audit defects turned into regression gates
  • 775 validated SQL objects deployed in dependency order with SHA-256 checksums via Azure DevOps
LLMsAzure AI FoundryEvals & guardrailsPythonCI/CD
Research · PhD

Deep RL agents that test route instructions

Reinforcement learning agents act as simulated users to check whether natural-language navigation instructions can actually be followed in complex buildings.

  • PPO and DQN agents (PyTorch, Stable-Baselines3, Gymnasium, TextWorld) with curriculum learning
  • 742 synthesised indoor environments; training runs up to 1M iterations on A100 GPU clusters
  • Evaluation harness with statistically significant gains over baselines; published in IJGIS and JOSIS
Reinforcement learningPyTorchNLPEvaluation
Code: Wayfinder-RL →
Data & MLOps · Parks Victoria

Production validation for an AI-assisted cloud migration

The engineering around the LLM pipeline: an evidence-based, restartable runbook that moved a corporate database to Azure PostgreSQL with production kept read-only throughout.

  • Restartable phases, preflight checks, target allow-listing, guarded rollback and distributed loading across machines
  • Source-vs-target validation of rows, columns, stable IDs, spatial references and geometry validity
  • Load-tested at 300 concurrent connections with zero failed connections
AzurePostgreSQLPythonAzure DevOpsTesting
Optimisation · University of Bonn

Optimal landmark placement in buildings

Multi-objective integer programming that places landmarks to reduce ambiguity in indoor navigation instructions.

  • MILP models solved with Gurobi, plus parallel and evolutionary (DEAP) approaches
  • Computational geometry and graph analysis of floorplans
  • Published in Computers, Environment and Urban Systems (2024)
GurobiMILPNetworkXPython
Code: Landmark_Placement →
RAG · Research

Location extraction from heritage reports

An LLM and retrieval pipeline that turns unstructured cultural heritage reports into structured, mapped records.

  • Chunking, embeddings and retrieval with LangChain and ChromaDB
  • Grounded extraction of places, dates and entities, geocoded and stored in PostGIS
  • Flask service for querying and visualisation
RAGLangChainChromaDBPostGIS
ML · Consulting

Forecasting and anomaly detection across 20+ ports

Distributed time-series forecasting and anomaly detection for sedimentation at DTCE Consulting Engineers.

  • PySpark on HPC, cutting detection time by 42% and sensor costs by over 65%
  • LSTM tide prediction integrated into field technicians' workflows
  • Parallel processing that cut terabyte-scale contour cleaning from 3+ months to one week
PySparkLSTMTime seriesHPC

Experience

Government, research and consulting: building data and ML systems that people rely on.

  1. Aug 2025 – Present

    Senior Spatial Analyst

    Parks Victoria · Melbourne
    • Helped build an LLM-assisted, evaluation-gated code translation pipeline (Microsoft Foundry models, GitHub Copilot CLI) that shipped 775 validated objects
    • Designed the Azure PostgreSQL migration runbook: validation, recovery, guarded rollback and an Azure DevOps release pipeline
    • Python tooling and large-scale data pipelines on a government data platform (100M+ records)
  2. Dec 2021 – Aug 2025

    PhD Researcher, Reinforcement Learning & Spatial AI

    University of Melbourne
    • Deep RL agents for validating route instructions; LLM/RAG pipelines for location extraction
    • Four peer-reviewed papers from the thesis; Melbourne Research Scholarship
  3. Feb 2023 – Jun 2025

    Tutor, Spatial Data Management and Spatial Data Analytics

    University of Melbourne
    • Postgraduate tutorials in Python data analysis, databases and statistics
  4. Nov – Dec 2023

    Visiting Researcher (Optimisation)

    University of Bonn, Germany
    • Integer programming models for landmark placement (Gurobi)
  5. Dec 2017 – Dec 2021

    Senior Spatial Data Scientist

    Darya Tarsim Consulting Engineers (DTCE)
    • Led a team of 5–7 data specialists within a 15-person engineering group
    • Took ML and time-series models from concept to production with engineers and domain experts
    • Distributed forecasting and anomaly detection (PySpark) cutting detection time by 42% and sensor costs by over 65%
  6. May 2014 – Dec 2017

    Spatial Data Scientist

    Darya Tarsim Consulting Engineers (DTCE)
    • LSTM tide forecasting, flood susceptibility modelling and parallel processing of terabyte-scale data

Skills

The tools I use most, grouped by what they're for.

AI & machine learning

LLMs & agentsRAGStructured outputsEvals & guardrailsAzure AI FoundryPyTorchStable-Baselines3Deep RL (PPO, DQN)scikit-learnLSTM / time seriesGurobi / MILPLangChainChromaDB

Engineering & MLOps

PythonEvaluation harnessesRegression gatespytestAzure DevOps CI/CDDockerKubernetesGitLinux & GPU clustersFlaskStreamlit

Data

SQLPySparkpandasNetworkXPostgreSQL / PostGISOracleAzurePower BIGeospatial data (GeoPandas)

Publications

Peer-reviewed research on reinforcement learning, optimisation and spatial decision-making.

Education

  • PhD, Infrastructure Engineering (Geomatics)University of Melbourne · 2021–2025 · Melbourne Research Scholarship
  • Master of Science, Geospatial Information SystemsUniversity of Tehran · 2012–2015
  • Bachelor of Civil Engineering – SurveyingUniversity of Tehran · 2008–2012 · ranked 1st nationwide in the entrance exam

Certifications

  • Reinforcement Learning SpecializationUniversity of Alberta / Coursera
  • Large Language Models BootcampNVIDIA
  • RAG with LangChainDataCamp
  • DevOps Professional CertificatePagerDuty · 2025
  • Docker FoundationsDocker, Inc.

Let's talk

Open to Applied AI, ML engineering and data science roles in Melbourne, hybrid or remote.

rasheibani@gmail.com LinkedIn GitHub