Parham Hamouni
Verified Expert in Engineering
Data Scientist and Developer
Toronto, ON, Canada
Toptal member since March 14, 2022
Parham is a senior LLM engineer and PhD researcher with 6+ years of shipping production AI systems across manufacturing, legal, and finance. At Toronto Metropolitan University, his PhD applies DPO and RLHF to temporal graph reasoning, achieving 83% pairwise accuracy and outperforming baselines by 12–15pp. In production, Parham works with RAG pipelines, fine-tuned transformers (FinBERT, Qwen, LLaMA-2), and full MLOps stacks. At Kraft Heinz, his RAG system cut troubleshooting time by 30%.
Portfolio
Experience
- Python - 8 years
- Deep Learning - 8 years
- Graphs - 8 years
- Natural Language Processing (NLP) - 8 years
- Artificial Intelligence (AI) - 8 years
- Statistics - 6 years
- Computer Vision - 6 years
- SQL - 6 years
Preferred Environment
Unix, Python, SQL, Deep Learning, Graphs, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP), Azure, Amazon Web Services (AWS), Pandas
The most amazing...
...result I've achieved improved the accuracy of a knowledge graph entity prediction problem from 35 to 95 percent using a simple yet effective technique.
Work Experience
AI Engineer
Adams Fence, LLC
- Built an end-to-end document intelligence pipeline for construction bid automation: ingested multi-page PDF/Bluebeam plan sets via agentic OCR and spatial layout analysis, eliminating manual data entry in RFP-to-quote workflows.
- Engineered a production-ready Streamlit application deployed on AWS, enabling estimators to process technical construction documents and extract structured project elements without manual review, reducing quoting time by an estimated 60%.
- Designed a layout-aware document parsing system using spatial context and page structure to accurately identify and extract domain-specific construction elements across heterogeneous multi-page technical plan formats.
Senior NLP Engineer
Fog City Code Works
- Led development of a paragraph-level text classification system for finance interview transcripts (10 years of labeled data, 30-40 categories).
- Implemented end-to-end pipeline: data cleaning, transformer fine-tuning (FinBERT/Qwen), experiment tracking (MLflow), and evaluation (macro-F1, calibration).
- Extended the pipeline with an agentic RAG system in phase two, incorporating multi-stage retrieval and re-ranking, improving micro-F1 by 15% over the fine-tuned baseline.
Senior Data Scientist
The Kraft Heinz Company
- Designed and deployed a retrieval-augmented generation (RAG) system using large language models (LLMs) to assist maintenance technicians, reducing troubleshooting time by 30%.
- Built a reactive maintenance prediction model that decreased unplanned downtime incidents by 25%.
- Developed an unsupervised object detection system for retail shelves using DINO, improving shelf recognition accuracy by 40%.
- Utilized a full MLOps stack: HuggingFace Transformers, FAISS vector store, LangChain, Docker, PyTorch, and MLflow for experiment tracking and model deployment on Azure.
Applied Research Scientist
Alexi
- Created synthetic QA pairs from Canadian legal documents to train a legal retrieval system, increasing retrieval accuracy by 10%.
- Developed a classifier to distinguish legal facts and principles in case law, achieving 92% classification accuracy and handling end-to-end pipeline development.
- Trained a self-supervised language model on the sentences that accurately assigned relative risk in the news to the companies, assuming that headers and paragraphs of 10-K forms have the same concepts.
Applied Research Scientist
Crater Labs Inc
- Trained a self-supervised language model on the sentences that accurately assigned relative risk in the news to the companies, using news about companies in the stock market and their relative 10-K forms in the SEC database.
- Improved edge pair prediction by more than 60 percent to 95 percent Hits@1 score by reframing a knowledge graph entity prediction problem as a sentence pair classification.
- Predicted the best price discount for a startup company with a regression problem approach using multi-relational graph embedding techniques to create relative features.
- Used question answering data scraped from the client's forum to create a dialog generation network based on GPT-2 with a persona context feature. Also applied spectral regularization, which improved the perplexity score.
- Handled manufacturing defect detection in highly imbalanced data. Using conditional GANs and image-to-image translation techniques augmented the data so the data imbalance was managed and the classifier overcame overfitting. Used Azure environment.
- Applied unsupervised image encoding techniques such as adversarial autoencoders and CNN models to find similar images in a proprietary photoshoot dataset. Used AWS environment.
Experience
TF-MF: Improving Multiview Representation for Twitter User Geolocation Prediction
https://ieeexplore.ieee.org/document/9073316Impact-DPO: Predicting Scholarly Influence with Temporal Preference Alignment
https://github.com/parhamouni/impact-dpoEducation
PhD Candidate in Electrical and Computer Engineering
Toronto Metropolitan University - Toronto, ON, Canada
Master of Science in Civil Engineering (Transportation Planning)
Concordia University - Montreal, QC, Canada
Skills
Libraries/APIs
Pandas, XGBoost, PyTorch
Tools
AWS Command Line Interface (CLI)
Languages
Python, SQL
Platforms
Unix, Azure, Amazon Web Services (AWS), Docker
Other
Deep Learning, Graphs, Natural Language Processing (NLP), Computer Vision, Statistics, Artificial Intelligence (AI), Generative Pre-trained Transformers (GPT), Large Language Models (LLMs), Retrieval-augmented Generation (RAG), LLM Fine-tuning, Reinforcement Learning, Machine Learning, MLflow, Unsupervised Learning, Hugging Face, Optical Character Recognition (OCR), LangChain, FAISS, Document AI/OCR, APIs, Reinforcement Learning from Human Feedback (RLHF), Direct Preference Optimization
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring