Hi, I'm Pegah Salehi

    Still learning, always building, occasionally sleeping.Open to meaningful collaborations.

    Applied AI Engineer|

    I build applied AI systems that move beyond prototypes, including RAG pipelines, generative AI workflows, and computer vision models.

    Portrait of Pegah Salehi

    AI & Research Snapshot

    5+
    Years Experience
    10+
    Projects Completed
    20
    Publications
    850
    Citations

    Technologies I Work With

    Tools and frameworks I use to build applied AI systems, generative AI solutions, and intelligent products.

    About Me

    Education

    UiT The Arctic University of Norway logo
    PhD
    University of Tromsø / UiT The Arctic University of Norway

    Doctoral research focused on applied AI, avatar-based training systems, generative visual models, and human-AI interaction.

    Download PhD Thesis
    Aug 2025 – May 2026

    Applied AI Engineer, Min Sone AS, Oslo

    AI product company developing interactive training and digital learning solutions

    - Built a RAG-based chatbot system for Krifa, integrating document retrieval, LLM-powered response generation, and a user-facing chat interface to improve access to company knowledge. - Contributed to the development of Ågo Course, a full-stack interactive training platform, focusing on Python backend development and AI components within a Next.js/React-based product stack. - Created AI-generated training videos with Google Veo and designed interactive course content for electrician training scenarios. - Configured a web-based community platform using WordPress, BuddyBoss, and Elementor to support digital learning product delivery.

    Oct 2021 – Oct 2025

    PhD Researcher, Simula Metropolitan Center for Digital Engineering (SimulaMet), Oslo

    Research center focused on applied AI, digital engineering, and human-AI interaction

    - Developed interactive AI-powered child avatar systems for investigative interview training by combining talking-head generation models, game-engine-based avatars, NLP, STT/TTS, and real-time interaction pipelines. - Optimized NeRF-based talking-head generation by integrating Whisper-based audio feature extraction, reducing processing time and improving low-latency avatar interaction. - Built emotionally expressive avatar prototypes using Unreal Engine 5, MetaHuman, NVIDIA Omniverse Audio2Face, and Live Link for audio-driven facial animation. - Designed and evaluated VR/avatar training prototypes using Unity 3D, Oculus Quest 2, and SALSA LipSync Suite, and analyzed user engagement using ANOVA, mixed-effects models, and SEM.

    Sep 2024 – May 2025

    AI Engineer (Freelance), Truviz Inc., Remote, U.S.

    A startup developing conversational AI and avatar-based visual interaction systems

    - Developed talking-avatar features for a conversational LLM application, enabling generated text and audio responses to be presented through expressive AI-driven visual agents. - Implemented, tested, and customized open-source GAN, diffusion, and talking-head generation models for avatar generation, facial animation, and visual response generation workflows.

    Core Capabilities

    Generative AI for Vision & Avatars

    Applying GANs, diffusion models, and NeRF-based methods to build visual AI systems, including talking-head avatars, facial animation pipelines, and generative workflows for medical image analysis.

    Medical Image Analysis

    Developing deep learning and generative AI methods for histopathology image analysis, stain normalization, synthetic data generation, and image segmentation tasks.

    Talking-Head & Avatar Systems

    Designing interactive AI avatar systems by combining talking-head generation, facial animation, STT/TTS, NLP, Unity/Unreal Engine, MetaHuman, and NVIDIA Audio2Face.

    LLM & RAG Applications

    Building LLM-powered applications with document retrieval, vector search, prompt engineering, LangChain, Azure OpenAI, and RAG architectures for knowledge-based assistants.

    Human-AI Training Systems

    Creating AI-driven training environments with virtual avatars, VR/3D interaction, conversational agents, and adaptive simulation workflows for professional learning.

    Applied AI Product Engineering

    Turning research prototypes into user-facing AI products by integrating models, APIs, cloud services, and interactive web platforms into practical workflows.

    Projects

    What I’ve built along the way, each project shaped my journey and taught me something new.

    Pix2Pix-based Stain-to-Stain Translation

    Pix2Pix-based Stain-to-Stain Translation

    Python
    cGAN
    Deep Learning
    Medical Image Analysis

    A deep learning project for robust stain normalization in histopathology images using conditional GANs and the Pix2Pix image-to-image translation framework.

    Problem
    Stain variation in H&E histopathology images caused by different scanners and laboratory protocols can reduce the reliability of automated medical image analysis systems.
    Solution
    Implemented a Stain-to-Stain Translation approach using the Pix2Pix framework with a U-Net generator and PatchGAN discriminator to learn stain normalization from paired histopathology image patches.
    Result
    The method produced visually consistent normalized H&E images while preserving tissue morphology, and achieved strong quantitative performance compared with traditional stain normalization methods.
    SinGAN-Seg: Synthetic Training Data Generation

    SinGAN-Seg: Synthetic Training Data Generation

    Python
    SinGAN
    Synthetic Data
    Medical Image Segmentation

    A synthetic data generation pipeline for medical image segmentation using SinGAN and single-image training.

    Problem
    Medical image segmentation often suffers from limited annotated data, especially in healthcare where data collection and labeling are expensive and time-consuming.
    Solution
    Built a SinGAN-based synthetic data generation pipeline that creates medical images and corresponding segmentation masks from a single training image, making it possible to expand small datasets.
    Result
    The synthetic data improved segmentation performance in low-data settings and showed that generated datasets can be useful alternatives or supplements to real annotated medical data.
    Whisper AFE for Real-Time Talking Portrait Synthesis

    Whisper AFE for Real-Time Talking Portrait Synthesis

    Python
    Whisper
    NeRF
    Talking Heads
    Real-Time Avatars

    A real-time talking portrait synthesis project that uses Whisper for faster audio feature extraction and more responsive AI avatar interactions.

    Problem
    Audio feature extraction in talking-head generation can introduce high latency, reducing responsiveness and realism in interactive avatar systems.
    Solution
    Integrated Whisper as an efficient AFE module and evaluated it with RAD-NeRF and ER-NeRF against DeepSpeech, Wav2Vec, and HuBERT across multiple datasets.
    Result
    Achieved reduced latency while preserving visual quality and improving lip-sync responsiveness for real-time talking portrait applications.
    Emotionally Expressive Child Avatars

    Emotionally Expressive Child Avatars

    Unreal Engine
    MetaHuman
    Audio2Face
    Interactive Avatars
    HCI

    A real-time emotionally expressive child avatar system using Unreal Engine MetaHuman and NVIDIA Audio2Face for training applications.

    Problem
    Virtual child avatars used in training systems often lack believable facial emotions, reducing realism, empathy, and user engagement in sensitive interview simulations.
    Solution
    Built a multimodal avatar pipeline combining Unreal Engine 5 MetaHuman, NVIDIA Omniverse Audio2Face, TTS, LLM-based dialogue, and Live Link for real-time facial animation.
    Result
    Evaluated the system in a user study with 70 participants, showing how audiovisual alignment, facial morphology, and voice consistency affect emotion recognition and perceived realism.
    2D & 3D VR Training for Child Interview Skills

    2D & 3D VR Training for Child Interview Skills

    Unity
    VR
    GPT-3
    User Study
    Training Systems

    An AI-driven child interview training platform comparing 2D desktop and 3D VR environments for investigative interview skill development.

    Problem
    Professionals need scalable and realistic ways to practice investigative interviews with children, but traditional actor-based training is costly, difficult to repeat, and limited in availability.
    Solution
    Developed a Unity-based training system integrating a fine-tuned GPT-3 dialogue model, STT/TTS speech interaction, child avatars, and both 2D desktop and Oculus Quest VR environments.
    Result
    A mixed-method user study showed that 3D VR improved immersion, flow, and virtual fidelity, while the 2D setup was preferred for usability and comfort.
    FoliaScan: Azure ML Crop Disease Classification

    FoliaScan: Azure ML Crop Disease Classification

    Python
    PyTorch
    Azure ML
    ResNet18
    MLflow
    Streamlit

    An end-to-end computer vision and MLOps project for tomato leaf disease classification, covering PyTorch ResNet18 training, Azure Machine Learning, and Streamlit cloud inference.

    Problem
    Image classification portfolio projects often stop at local model training, leaving gaps in reproducibility, cloud orchestration, deployment validation, and usable inference workflows.
    Solution
    Built a leakage-safe PlantVillage Tomato pipeline with PyTorch and ResNet18, Azure ML training and MLflow tracking, model registration, Managed Online Endpoint validation, a Streamlit inference client, and GitHub Actions CI.
    Result
    Reached 92.07% final test accuracy with a macro F1 score of 0.89665 and validated real end-to-end Streamlit-to-Azure inference before removing the cloud endpoint to avoid ongoing Azure costs.
    Advanced RAG Knowledge Assistant
    In Progress

    Advanced RAG Knowledge Assistant

    RAG
    Azure OpenAI
    Vector Search
    LangChain
    AI Assistant

    An advanced RAG-powered knowledge assistant for reliable document-based question answering with citations, semantic search, and LLM response generation.

    Problem
    Organizations often store important knowledge across PDFs, internal documents, and fragmented resources, making it difficult for users to quickly find accurate and trustworthy answers.
    Solution
    Designed a retrieval-augmented generation pipeline using document ingestion, chunking, embeddings, hybrid/vector search, semantic reranking, prompt engineering, and grounded LLM responses with source citations.
    Result
    The system is being developed to improve knowledge access, reduce manual search time, and provide accurate AI-assisted answers for internal users and training workflows.

    Micro Projects

    • Anar

      A personal nutrition tracker for saving everyday meals, logging them quickly, and following calories and macros over time.

      Next.js
      TypeScript
      Tailwind CSS
      Supabase
      PostgreSQL
      PWA
    • AI Portfolio with RAG Chat

      Personal portfolio with a RAG chat assistant for exploring my projects, publications, experience, and skills through interactive AI conversation.

      React
      TypeScript
      Vite
      Tailwind CSS
      Node.js
    • Min Sone Website

      AI-focused business website built to present digital services, product vision, and user-facing information through a clean and modern web experience.

      WordPress
      Elementor
    • Community Platform

      Configured a web-based community platform using WordPress, BuddyBoss, and Elementor to support digital learning product delivery.

      WordPress
      BuddyBoss
      Elementor

    Publications

    Where my experiments turned into papers, from medical images and generative models to AI avatars that can talk back.

    20
    Publications
    850
    Citations
    12
    h-index
    13
    i10-index

    Contact

    Always up for collaborations and interesting projects.

    Let's build something intelligent

    Pegah logo

    © 2026 Pegah Salehi. All rights reserved.

    Built with and