A. Yousefiramandi

Machine Learning · NLP · Applied Research Barcelona, Spain

Amirhossein
Yousefiramandi

I own the ranking models, data pipelines and evaluation framework behind clinical-trial site tiering at IQVIA — and publish on patent-domain NLP, LLM fine-tuning and representation learning.

LinkedIn GitHub Google Scholar CV (PDF)

6+ Years in ML/AI Since 2020
6 arXiv preprints 4 first-author · 2 sole 3 Patents 1 sole inventor 14 Open-source projects All public

About

Six-plus years of production machine learning, and the research that keeps coming out of it.

Barcelona, Spain

Six-plus years building and deploying large-scale ML systems, now at IQVIA in Barcelona. The site-tiering work is a family of ranking models that score and tier investigator sites so clinical-trial teams get an evidence-backed shortlist for study placement — plus the evaluation framework, orchestration and monitoring that decide when a new model is actually good enough to replace the one in production.

Before that, nearly three years at Clarivate on patent-domain NLP and generative AI: pretraining domain-specific language models, building taxonomy-guided classifiers, and productionising multi-agent prior-art research systems on AWS.

The work runs end to end — data pipelines, pretraining and fine-tuning in PyTorch and Spark/Databricks, and low-latency serving on AWS (SageMaker, Fargate, Bedrock).

Publications

Six preprints, four of them first-author, two written alone.

  1. When Does Synthetic Patent Data Help? Volume–Fidelity Trade-offs in Low-Resource Multi-Label Classification

    First author · with C. Cooney ·

    arXiv:2605.24297First author

  2. Benchmarking Patent Embeddings: A Multi-Task Evaluation of 22 Models Across Retrieval, Classification, and Clustering

    First author · with C. Cooney ·

    arXiv:2605.24296First author

  3. Learning Selective Merge Policies for Deadline-Constrained Coded Caching via Deep Reinforcement Learning

    Sole author ·

    arXiv:2605.15236Sole author

  4. Cumulative-Goodness Free-Riding in Forward-Forward Networks: Real, Repairable, but Not Accuracy-Dominant

    Sole author ·

    arXiv:2605.06240Sole author

  5. Fine-Tuning Causal LLMs for Text Classification: Embedding-Based vs. Instruction-Based Approaches

    First author · with C. Cooney ·

    arXiv:2512.12677First author

  6. Patent Language Model Pretraining with ModernBERT

    First author · with C. Cooney ·

    arXiv:2509.14926First author

Patents

Three filings in applied machine learning; sole inventor on one.

  1. Prior Art Search Using Artificial Intelligence Agents

    Sole inventor · Camelot UK Bidco Ltd. / Clarivate

    Docket I03.00450000PendingNo public link

  2. Training an Artificial Intelligence Classifier Using LLM-Labeled Training Data

    First inventor

    U.S. App. 19/029,628US20260212116A1Published

  3. Hybrid Artificial Intelligence Classifier

    Co-inventor

    U.S. App. 18/590,188US20250272608A1Published

Selected Experience

Seven positions, newest first — clinical trials, patents, search, speech and vision.

2020–present


  1. – Present

    IQVIA

    Senior Machine Learning Engineer

    Barcelona, Spain

    • Own the ranking models and the data pipeline behind clinical-trial site tiering end to end — the system trial teams use to choose investigator sites for studies. The model work is data-centric: improve what the two-tower ranker is fed, not the architecture.
    • Build and operate the evaluation framework that decides whether a candidate model replaces the one in production, ranking each against the incumbent on versioned datasets. Guard the measurement itself, not only the model: drift and feature-space-collapse detection, plausibility ceilings on predictions, de-duplication of identical retrains. An evaluation run that reports nothing alerts like a failed one.
    • Moved nightly ingestion off cron-offset scheduling onto orchestrated Airflow DAGs with real dependencies, retries, data-quality gates and end-to-end failure alerting. The nightly window fell from ~9 h to ~5 h.


  2. – 

    Clarivate

    Senior Data Scientist, Applied Science

    Barcelona, Spain

    • Designed and shipped the ML core of the Derwent AI Classifier, which mapped a customer's patents onto their own taxonomy: hierarchical classifiers trained on demand where labels existed, LLM labeling where they did not. $10M+ in incremental revenue and a 30% lift in renewal contracts; two published USPTO applications cover the work.
    • Replaced a manual examiner workflow with a multi-agent prior-art and novelty research system on AWS, with guardrails against prompt injection and data leakage, and built the per-release evaluation suite that made its improvements measurable and customer-reportable. Sole inventor on the patent filing that covers it.
    • Pretrained patent-domain masked language models from scratch on a curated 60M+ patent corpus with a custom tokenizer. The resulting model beat general ModernBERT on three of four internal benchmarks at over three times PatentBERT's inference throughput.


  3. – 

    Double Avatars

    AI Engineer

    Spain

    • Virtual try-on and 3D avatarization models for e-commerce.


  4. – 

    Freelance

    Freelance Data Scientist

    Spain

    • Semantic search improvements (MRR@15 +8%) and multimodal price-estimation and recommender systems (MAE −12%).


  5. – 

    Semrush

    Data Scientist

    Spain

    • Active-learning NLP pipelines on large-scale web text.


  6. – 

    CafeBazaar

    Data Scientist

    Tehran, Iran

    • Semantic app search (Recall@20 +9%); POI extraction with YOLO / OCR / SATRN (F1 53% → 65%, half the inference time); multi-GPU segmentation training (+11% mIoU).


  7. – 

    Baharan

    Data Scientist

    Iran

    • Facial analysis and recognition models for surveillance systems.

Open-Source Projects

Fourteen public repositories across speech, vision, language and agents.

  1. SSM Benchmark: Mamba-2 vs. Transformer vs. Jamba

    Apples-to-apples training and inference benchmark on a ~1.5B-token FineWeb-Edu slice; MQAR, induction copy and selective copy; flat decode latency for Mamba against Transformer KV growth.

  2. Text-JEPA Experiments & AF-JEPA

    JEPA target-design studies on WikiText-2 and FineWeb-Edu (PyTorch, MiniLM, Qwen 2.5-7B); raw-target top-1 retrieval 0.242 vs. semantic cards 0.183.

  3. Multilingual LLM Adaptation (Swahili Biomedical)

    MAD-X-style stacked adapters on a frozen AfroXLMR-large backbone, evaluated on MMLU-ProX-Swahili clinical and MasakhaNER.

  4. Instruction-Tuned LLM for Text Classification

    LoRA / QLoRA fine-tuning with TRL and PEFT, reaching ~94% accuracy.

  5. Multi-Modal Hateful Content Classification

    CLIP / SigLIP late fusion with multi-task learning on MMHS150K (macro-F1 0.569, ROC-AUC 0.783).

  6. VLM Fine-Tuning via JSON Generation

    Qwen2-VL, PaliGemma2 and SmolVLM for multi-label detection (Qwen2-VL + LoRA, micro-F1 0.611).

  7. Multi-Label Emotion Classification

    RoBERTa-based classifier over 27 emotions (micro-F1 0.528, ROC-AUC 0.952).

  8. Whisper Multi-Task ASR + Voice Intensity

    Joint ASR and loudness regression (WER 4.70%, intensity RMSE 0.73); Gradio demo and SageMaker deployment.

  9. Wav2Vec2 Multi-Task: CTC ASR + Intensity

    Wav2Vec2-base CTC with an intensity head and attention-masked mean pooling (val WER ≈12.9%).

  10. CTC ASR Baseline (Speech-to-Text)

    End-to-end Wav2Vec2 pipeline on Common Voice 17.0, with ONNX export, CLI inference and SageMaker autoscaling / HPO.

  11. Image–Text Retrieval (CLIP / SigLIP)

    Flickr30k R@1 92.3% image→text and 79.0% text→image, with FAISS-ready embeddings.

  12. Keyword Spotting (Wav2Vec2, ONNX, SageMaker)

    Real-time streaming inference on Speech Commands v2, with realtime, serverless and batch deployment helpers.

  13. Emotion-Aware TTS Style Transfer

    SpeechT5, WavLM, ECAPA and HiFi-GAN combined through StyleAdaptor / StyleSpeakerFusion, with a Gradio UI.

  14. Any-to-Any Speech Conversion

    SpeechT5-VC with ECAPA and HiFi-GAN for voice conversion from a short reference clip; fine-tuned on CMU ARCTIC.

View all projects on GitHub

Technical Glossary

The working vocabulary — what is actually used, not what is merely recognised.

6 fields · 3 spoken languages
Research & Modeling
  • Transformer architectures (MHA, GQA, MoE)
  • attention variants (Flash, paged)
  • positional encodings (RoPE, ALiBi)
  • state-space models (Mamba, Mamba-2, Jamba)
  • RLHF / DPO / GRPO
  • knowledge distillation
  • continued pretraining
  • controlled ablation design
  • two-tower retrieval / learning-to-rank
  • drift and model-quality evaluation
Frameworks
  • PyTorch
  • TensorFlow
  • Hugging Face (Transformers, TRL, PEFT, Diffusers)
  • MosaicML Composer
  • MLflow
  • Scikit-learn
  • LightGBM
LLMs & Agents
  • LoRA / QLoRA
  • UnSloth
  • DSPy
  • DeepAgents
  • LangGraph
  • LlamaIndex
  • MCP
  • Strands Agents
  • AgentCore
  • Google ADK
  • DeepEval
Cloud & MLOps
  • AWS (SageMaker, Glue, Lambda, Fargate, Bedrock, Step Functions)
  • Azure AI
  • Databricks / Spark
  • Airflow
  • MLflow
  • Datadog
  • Terraform
  • Kubernetes
  • Docker
  • Jenkins
  • ONNX
Languages & Tools
  • Python
  • C++
  • Bash
  • SQL
  • Git
  • LaTeX
Quantitative Research
  • Mathematical optimization
  • mathematical modeling
  • statistical learning
  • information theory
Spoken
  • English (fluent)
  • Spanish (working)
  • Persian (native)

Distinctions

Four things worth stating plainly.

  • Promoted at Clarivate within less than a year for exceptional performance.

  • Three patents from the Clarivate work, sole inventor on one — alongside six arXiv preprints, two of them sole-authored.

  • Twice promoted at CafeBazaar within one year for outstanding performance.

  • Ranked 16th nationwide in Iran’s master’s-programme entrance examination.

Education

Two degrees in electrical engineering; the second in artificial intelligence.

2014–2021

  1. –

    Sharif University of Technology

    MSc, Electrical Engineering (Artificial Intelligence)

    Tehran, Iran

    University site

  2. –

    Isfahan University of Technology

    BSc, Electrical Engineering

    Isfahan, Iran

    University site