AI Perception Engineer/Autonomous Driving

Mahdi
Farhadi

I turn research-grade perception models into systems that actually drive a car — 3D object detection, LiDAR–camera fusion, and real-time perception for Level-4 autonomy.

View projects Download résumé Get in touch
  • LOC Tehran, Iran
  • EXP 4+ yrs · 3D vision stack
  • STATUS Open to PhD positions
SCROLL

01 About

Perception that survives the messy gap between benchmark and field.

I'm an AI developer who likes turning research-grade perception models into systems that actually drive a car. Over the past four years I've worked across the full 3D-vision stack — designing, training, pruning, and deploying detectors like PointPillars and BEVFusion for real-world ADAS.

Today I help design a Level-4 self-driving system for Iran's all-electric YOOZ vehicle, build digital-twin talking-head avatars, and run real-time inference on embedded and ROS 2 platforms. What I enjoy most is the messy gap between a model that shines on a benchmark and one that holds up in the field — and as a PhD candidate I'm channelling that experience into research on perception for autonomous driving.

  • 0 Years across the
    3D-vision stack
  • 0 BEVFusion releases
    shipped to SoC teams
  • 0 Detection accuracy
    gained on in-house data
  • 0 MSc GPA
    out of 20

RESEARCH INTERESTS

  • Multi-modal Sensor Fusion
  • 3D Object Detection & Tracking
  • Real-time Perception
  • Semantic Segmentation
  • Foundation Models
  • Vision Transformers
  • Self-supervised Learning
  • Edge Optimization
  • Robust Perception
  • Uncertainty Quantification
  • Generative AI

02 Experience

Where the work happened.

  1. 2025 — Present

    AI Developer Mid Level

    MCI (NAK) — R&I Department · Tehran, Iran

    • Design and help manage a Level-4 ADAS self-driving system for the Iranian all-electric YOOZ car.
    • Build and integrate the autonomy stack — perception, sensor fusion, and real-time decision-making toward full self-driving.
    • Develop a digital twin of a talking-head avatar product, connecting generative AI with real-time simulation.
  2. 2024 — Present

    Algorithm Developer Senior Level

    Software Motion · Suzhou, China / Iran

    • Promoted to senior after joining mid-level; work in the perception team on PointPillars and BEVFusion for L2/L3 ADAS.
    • Build full training pipelines on in-house datasets — preprocessing, annotation, training from scratch, and evaluation.
    • Lead model pruning & optimization, compressing detectors for real-time inference on constrained hardware.
    • Validate perception online inside a ROS 2 pipeline; own model releases and the team Git repository for on-vehicle deployment.
    • Contributed to collaborations with major Chinese automotive companies including Tucson, JAC, and AICC.
  3. Mar — Jul 2024

    Data Scientist

    Nerd Studio AI · Milwaukee, WI, USA

    • Worked on large generative models; developed the Indox framework integrating OpenAI and Mistral models.
    • Packaged Indox for PyPI for easy developer access — github.com/osllmai/inDox.
  4. 2023

    AI Engineer

    IRIB · Tehran, Iran

    • Developed 3D-reconstruction models for multi-task use — body/face recognition and 3D object reconstruction.
    • Designed a user-friendly GUI to present the models as commercial products; project subsidized by the Iran National Elite Foundation.
  5. 2022 — Present

    Researcher

    IUST Machine Vision Lab · Tehran, Iran

    • Developed an original method to detect urban road obstacles in segmentation — led to a conference paper.
    • Object detection for crack identification in X-ray images (non-destructive testing); assessed the NAS-Unet model across datasets.
  6. Feb — Aug 2023

    Teaching Assistant

    Iran University of Science & Technology · Tehran, Iran

    • Taught the Machine Vision course (50 undergraduates) on Python and running GitHub code; administered and graded assessments.

03 Selected Work

Four systems, from raw sensors to real-time output.

P-01

Multi-Modal 3D Object Detection — BEVFusion

LiDAR + Camera Fusion · Software Motion · 2024 — Present

Detectors that fuse LiDAR and camera in a shared bird's-eye-view space. I boosted accuracy by roughly 10% through better data augmentation, view transformation, and fusion, then shipped nine versions — each one pruning the network, cutting latency, and sharpening the detection head. I also grew the backbone into a multi-task model, integrating lane and occupancy heads and MapTR-Head-V2.

  • BEVFusion
  • LiDAR–Camera
  • Occupancy Head
  • Lane / MapTR-v2
  • Pruning
  • MMDetection3D
FIG.01 Surround-camera 3D boxes · LiDAR BEV point cloud · predicted lanes · occupancy grid
FIG.02 A denser urban scene — fusion holds up where a single modality drops out
ROS 2 · LIVE PERCEPTION
P-02

Online Perception in ROS 2 — PointPillars

Real-time LiDAR detection · On-vehicle validation

The same detector, now streaming. I validate real-time perception online inside a ROS 2 pipeline — PointPillars consuming live LiDAR and publishing 3D detections frame by frame, so the model is judged the way it runs on the car, not just on a saved benchmark.

  • ROS 2
  • PointPillars
  • Real-time
  • TensorRT
  • On-vehicle

▶ Clip plays automatically — tap to pause. Recorded from the live visualiser.

FIG.03 Predicted link volumes across the network at four times of day
P-03

Traffic Forecasting from OD with a GRU Network

Spatio-temporal demand modelling · Digital-twin research

Forecasting how a whole city breathes. A GRU recurrent network takes origin–destination demand and predicts traffic volume on every link of the road network through the day — the morning surge, the midday plateau, the evening peak, and the quiet of 23:00 — feeding urban-traffic and digital-twin simulation.

  • GRU / RNN
  • Origin–Destination
  • Spatio-temporal
  • Traffic Simulation
  • Digital Twin
P-04

Road Anomaly & Obstacle Segmentation

MSc Thesis · IUST · Published at ICCKE 2023

My master's thesis: teaching a self-driving car to flag what doesn't belong on the road — a maintenance crew, livestock wandering across a mountain pass, debris the training set never showed. I built a segmentation method for out-of-distribution obstacles, which became Standardized ReACT Logits, published at ICCKE 2023.

  • Anomaly Segmentation
  • Out-of-Distribution
  • Deep Neural Nets
  • Autonomous Driving
Maintenance crew on the carriageway
Livestock on a mountain pass

ALSO Other things I've built

L4 Autonomy & Talking-Head Avatar

MCI (NAK) · 2025 — Present

Defined an L4 YOOZ ADAS project end to end, and deployed a SoulX-FlashHead API powering a conversational avatar chatbot.

VVC-flashhead

Indox — LLM Framework

Nerd Studio AI · 2024

A retrieval-and-generation framework integrating OpenAI and Mistral models, packaged and published to PyPI.

osllmai/inDox

3D Reconstruction Software

IUST · Elite Foundation · 2023

Body/face recognition and 3D object reconstruction, wrapped in a GUI for commercial use. Supervised by Dr. Amirkhani.

Line Navigation Robot

IEEE Student Branch · 2018

A line-following robot — circuit design in Proteus, microcontroller firmware in CodeVision — for the IEEE robotics course.

04 Toolkit

The stack I reach for.

PROGRAMMING

  • Python expert
  • C
  • MATLAB
  • SQL
  • Bash / Linux

CORE ML

  • PyTorch
  • Git expert
  • TensorFlow
  • Scikit-Learn
  • Docker
  • Deep Learning

VISION & 3D

  • BEVFusion
  • PointPillars
  • MMDetection3D
  • LiDAR–Camera Fusion
  • Semantic / Anomaly Seg.
  • YOLO family

DEPLOYMENT

  • ONNX
  • TensorRT
  • TF-Lite
  • CUDA
  • Quantization
  • ROS 2
  • FastAPI

GENERATIVE AI

  • LLMs · OpenAI
  • LangChain
  • HuggingFace
  • LlamaIndex
  • Text-to-Speech
  • Talking-Head Avatars

LANGUAGES

  • English professional
  • Persian native

05 Research & Education

Papers, degrees, and the people who vouch for me.

PUBLICATION · ICCKE 2023

Standardized ReACT Logits: An Effective Approach for Anomaly Segmentation in Self-driving Cars

M. Farhadi, S. M. Hazavei, S. B. Shokouhi

13th International Conference on Computer & Knowledge Engineering (ICCKE), Mashhad, Iran · pp. 312–316

DOI 10.1109/ICCKE60553.2023.10326295

MSc, Digital Systems Electronics

2021 — 2025

Iran University of Science & Technology · GPA 18.33/20

Thesis — road anomaly & obstacle segmentation with deep neural networks. Supervisors: Dr. S. B. Shokouhi, Dr. A. Amirkhani.

BSc, Electronic Engineering

2015 — 2020

University of Isfahan · GPA 15.06/20

Active member of the Robotics Society — designed line-following robots and took part in IEEE robotics activities.

ACHIEVEMENTS

  • Best Employee of the Month — Software Motion perception team DEC 2025
  • Council Membership — Robotics Society, University of Isfahan 2017–19

Let's talk

Building perception for autonomous systems —
and looking for a PhD to push it further.

Research collaborations, PhD positions, or a hard perception problem you can't get to hold up in the field — I'd like to hear about it.

farhadi76m@gmail.com Download résumé