Skip to content
View SajjadPSavoji's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report SajjadPSavoji

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
SajjadPSavoji/README.md

Sajjad Pakdamansavoji, Lead AI Research Scientist at Upwork. I post-train foundation models and keep them improving after they ship: today the LLM behind Upwork's job-post agent, before that robot VLAs and world models at Huawei.

Website Google Scholar Hugging Face LinkedIn X Forall Labs

Highlights: 1st place in the NeurIPS 2026 BEHAVIOR Challenge with team The North Star. 1M+ daily users on the LLM my team post-trains at Upwork. 13 papers and preprints on LLM post-training and evaluation, robot VLAs and world models, and 6D pose estimation. 3 patents filed on robot VLAs, world models and 3D vision.

About

I'm an AI Research Scientist with 6+ years of experience training, post-training and evaluating foundation models, from LLM agents and VLMs to robot VLAs and world models. I lead teams that take models from data to deployment, and my focus now is multimodal agents that learn continually from real-world feedback.

Currently, as Lead AI Research Scientist at Upwork, I lead a five-person team that owns the job-post generation model behind Upwork's AI agent, serving over 1M daily users. We replaced a paid API model in production with a post-trained Qwen3 LLM that is cheaper and rated higher, and I drive continual learning for the agent through a data flywheel built on gradient matching, with LLM-as-a-judge evaluation and online A/B tests.

Before that, at Huawei Noah's Ark Lab, I worked on embodied AI. I post-trained robot VLAs (π0.5, GR00T, OpenVLA) and DiT world models, built diffusion-based synthetic data engines for robot learning, and developed zero-shot 6D/9DoF pose estimation for warehouse robots. Our team, The North Star, won 1st place in the NeurIPS 2026 BEHAVIOR Challenge after placing 4th in 2025.

At the Vector Institute, I turned LLM and VLM research into agentic systems. I architected multi-step reasoning workflows and RAG pipelines that improved retrieval accuracy and safety, and led workshops on fine-tuning with LoRA and high-throughput serving with vLLM.

At Trans-Plan, I built 3D perception systems for urban environments. I led distributed fine-tuning of ViT-based foundation models for open-vocabulary detection, segmentation and tracking, raising tracking accuracy (MOTA) by 28% and reaching 92% accuracy in trajectory classification. I also built a real-time 3D visualization GUI for human-in-the-loop model refinement.

My early work at AVIR was a multimodal translation system for Iranian Sign Language: a MediaPipe and attention-based Bi-LSTM pipeline that translates video to text at over 30 FPS, fast enough for real-world use.

Awards: VISTA Scholarship, Vector Scholarship in AI, Lassonde and Founder's Entrance Scholarships, IEEE ITSC Best Application Paper Award, and York University's Academic Excellence Award.

Read my CV (PDF)

Featured open source

CALVIN-MAX: 7,828 cases of mid-task changes across 10 robot policies. Pooled task success: 69.9% with no change, 60.5% when the change is already in place, 63.0% when it arrives mid-task. OpenCaP: 21 open-weight LLMs writing robot code in 7,560 trials. Removing high-level helpers raises frame errors from 4% to 47% and API misuse from 26% to 60% of failing programs.

AutoPolish: one editing round. An image goes to Qwen2-VL persona critics, FLUX Kontext edits it, and the edit is kept only if the score rises; then the next round starts. Awesome Weight-Space Programming: the list grew from 31 to 606 entries between Nov 4 and Nov 15, curated by an agent.

MiniVLMDocEval: mean score over five document benchmarks. Qwen3.5-0.8B 70.3, InternVL3-1B 63.3, LFM2.5-VL-450M 59.9, LLaVA-OV-0.5B 52.8, SmolVLM2-500M 49.4, FastVLM-0.5B 36.8. colabctl: start, check and stop Google Colab GPU runtimes from the shell with colabctl up, colabctl status and colabctl down.

Selected publications

Venue Paper Links
Preprint OLGA: Online Label-Free Gradient Matching for Continual Language Model Training
First author · LLM continual learning
arXiv
Preprint Prompt Lottery in LLM-as-a-Judge: Measuring and Mitigating Sensitivity to Syntactic Rephrasing
LLM evaluation
arXiv
Preprint Fragility Is Not Reducible to Hardness: Conditional Causal Difficulty for LLMs
LLM robustness
arXiv
IROS 2026 How VLAs (Really) Work In Open-World Environments
Robot VLAs
arXiv · Page
ICRA 2026 Improving Robotic Manipulation Robustness via NICE Scene Surgery
First author · Robot manipulation
arXiv · Page
ICRA 2026 Distracted Robot: How Visual Clutter Undermine Robotic Manipulation
Robot manipulation
arXiv · Page
Preprint Do World Action Models Generalize Better than VLAs? A Robustness Study
World models · Robot VLAs
arXiv · Page
WACV 2026 WALDO: Where Unseen Model-based 6D Pose Estimation Meets Occlusion
First author · 6D pose
arXiv · Page
ECCV 2026 W AnyBox: Efficient Zero-Shot 9DoF Pose Estimation of Boxes for Robotic Manipulation
Equal contribution · 9DoF pose
arXiv · Page
ITSC 2024 Ground Plane Projection for Improved Traffic Analytics at Intersections
First author · 🏆 Best Application Paper
arXiv · Page

Full list on Google Scholar.

Experience

Experience: Lead AI Research Scientist at Upwork (2026 to now); Senior AI Researcher at Huawei Noah's Ark Lab (2024 to 2026); Machine Learning Engineer at Vector Institute (2023 to 2024, trained 40+ engineers); Machine Learning Engineer at Trans-Plan (2021 to 2023, 10+ TB ETL pipelines).

Education: M.Sc. Computer Science, York University · B.Sc. Electrical Engineering, University of Tehran

Toolbox

Toolbox. Training: PyTorch, JAX, DeepSpeed, FSDP, TRL, PEFT, LoRA/QLoRA, SFT, DPO, GRPO. Serving and agents: vLLM, SGLang, TensorRT, Triton, LangGraph, MCP, RAG, LLM-as-a-judge. Robotics: Isaac Sim, MuJoCo, ManiSkill, OmniGibson, LeRobot, CALVIN, ROS 2. Vision: SAM2, DINOv2, Diffusers, Open3D, PyTorch3D, OpenCV, OpenGL. Infra: Python, C++, AWS, Slurm, Kubernetes, Ray, Docker, Weights and Biases.

Recent

  • Nov 2026 · Released Awesome Weight-Space Programming, a daily, agent-curated map of LLMs that read, edit and write network weights
  • Nov 2026 · Released OpenCaP and CALVIN-MAX, two open Forall Labs studies of robot code generation and mid-execution task changes
  • Nov 2026 · OLGA, online label-free gradient matching for continual LM training, is on arXiv
  • Oct 2026 · 1st place in the NeurIPS 2026 BEHAVIOR Challenge with team The North Star

Say hello on LinkedIn or through my website. More work at sajjadpsavoji.github.io.
Icons: Lucide (ISC), Simple Icons and SVG Logos (CC0); brand marks belong to their owners.

Pinned Loading

  1. forall-labs/CALVIN-MAX forall-labs/CALVIN-MAX Public

    CALVIN-MAX: when does a mid-execution change hurt a robot policy? A 7,828-case benchmark on CALVIN, compatible with LIBERO-MAX.

    Python

  2. forall-labs/OpenCaP forall-labs/OpenCaP Public

    How open-weight models fail at writing robot code, and the smallest fix for each failure.

    Python

  3. forall-labs/MiniVLMDocEval forall-labs/MiniVLMDocEval Public

    Python

  4. forall-labs/AutoPolish forall-labs/AutoPolish Public

    AutoPolish: Improving Visual Aesthetics with Synthetic Audience Critique

    Python

  5. awesome-weight-space-programming awesome-weight-space-programming Public

    A curated list of weight-space programming: LLMs that read, edit and write the weights of other neural networks directly. Updated daily.

    Python 4 2

  6. ElderLab-York-University/Transplan ElderLab-York-University/Transplan Public

    my implementations for the transplan project

    Python