Ashkan Ansarifard

Ashkan Ansarifard

AI Research Software Engineer · PhD Student · Rome, Italy
GitHubLinkedInGoogle Scholar
Experience
Oct 2025 — Sep 2026
Research Collaborator, Sapienza University of Rome, DIAG
  • Built a differentiable logic-loss pipeline in PyTorch that injects LTLf constraints into autoregressive policies (Decision Transformer, Trajectory Transformer) via Gumbel-Softmax sampling on a DeepDFA backend.
  • Built reproducible, config-driven training infrastructure (checkpointing, resumable runs, Weights & Biases / rliable reporting) and evaluated on the ColourBomb safety benchmark.
  • Paper accepted at the SKILLED-LLMs 2026 workshop.
Jan 2025 — Present
AI Research Software Engineer, Amevista
  • Designed and shipped a queue-driven cloud native platform that runs large-scale batch LLM inference over Amevista's eyewear catalogue.
  • Built agentic automation for third-party review handling
  • Shipped React internal tools for lens-engraving personalisation and branded eyewear product-image preparation, used across design, marketing and production.
  • Built an AWS Bedrock pipeline to extract structured prescription data from scanned optical documents.
  • Led R&D&I initiative to deploy an open-weight LLM (GPT-OSS) on-premise, evaluating self-hosted inference as an alternative to cloud-based APIs for data privacy reasons.
  • Designed and benchmarked the serving infrastructure (vLLM/Ollama on GPU hardware), assessing trade-offs in output quality, throughput, and cost versus commercial LLM APIs.
  • AI Chatbot: Built an LLM-powered product-search assistant with a custom storefront chat widget, turning natural-language queries into product results.
Jan 2018 — Dec 2024
Software Engineer, ESTDCo
  • Built backend services, data-processing engines, real-time dashboards and reporting tools for RETINA Simulation, a black-oil reservoir simulator, and its surrounding engineering and maintenance systems.
  • Implemented database schemas, workflow automation, pytest suites and CI/CD pipelines, improving reliability and deployment cadence.
  • R&D&I initiative to extend RETINA's solver with data-driven surrogate modelling, Fourier Neural Operators in PyTorch, trained on simulation outputs to approximate pressure and saturation fields and accelerate physics runs.
  • Built the data pipeline and training/evaluation harness for the surrogate models: dataset versioning, reproducible experiments, error metrics against the reference solver, and packaging the trained models for the C++/JNI runtime.
Education
2026–present
PhD in Artificial Intelligence, Sapienza University of Rome
2025
Master of Science — Engineering in Computer Science, Sapienza University of Rome
2018
Bachelor of Science — Petroleum Engineering, Sharif University of Technology
Skills
Programming Languages
PythonTypeScriptJavaPHPC++SQL
ML & Deep Learning
PyTorchTensorFlowJAXHugging Face Transformersscikit-learnNumPy
LLMs & Generative AI
AWS BedrockLangChainLlamaIndexRAGVector databasesLLM evaluation (RAGASLLM-as-judge)LoRA / QLoRA fine-tuning
Reinforcement Learning & Planning
Offline RL (OSRL / DSRL)Decision Transformers
Model Serving & Optimization
NVIDIA TritonONNX RuntimeTensorRTvLLMQuantization
MLOps & Experimentation
Weights & BiasesMLflowDVCAirflowModel registries
Backend & Data
FastAPIFlaskREST APIsgRPCPostgreSQLMySQLMongoDBRedis
Cloud & Infrastructure
AWSGCP Vertex AIDockerKubernetesTerraform
Software Engineering
System designDistributed training
Languages
English C2, full professional
Italian B2, upper intermediate
Persian (Farsi) C2, native
German A1, elementary
References
Prof. Fabio Patrizi
Associate Professor, Department of Computer, Control, and Management Engineering (DIAG), Sapienza University of Rome, Italy
Sapienza pageGoogle Scholar
Prof. Elena Umili
Assistant Professor, Department of Computer, Control, and Management Engineering (DIAG), Sapienza University of Rome, Italy
Sapienza pageGoogle Scholar
© 2026 Ashkan Ansarifard