Welcome to my PhD Journey!
Milestones
IAAP–CO 2026 Student Scholarship Award
Selected by the Iranian-American Academics & Professionals (IAAP) Scholarship Selection Committee to receive the IAAP–CO 2026 Student Scholarship Award — the Land of Our Fathers Scholarship, sponsored by Dr. J. Rostami — recognizing academic achievement and contributions to the community.
Reflections from CVPR 2026
I attended CVPR 2026 in Denver this June. Three themes stood out.
Efficiency and on-device AI took center stage. The field has clearly pivoted from building capable models to running them on real hardware. Quantization, model compression, and hardware-aware optimization were everywhere, with Apple and NVIDIA both pushing on-device deployment and acceleration. It was encouraging to see efficiency treated as a first-class research problem.
Diffusion models have become a general-purpose tool. Beyond image generation, diffusion now powers classification, 3D reconstruction, and world models, with a strong focus on making inference fast enough for real-time and edge use.
World models and physics-aware vision are the next frontier. The growing consensus is that models must understand physical and causal structure, not just predict pixels, pointing toward vision systems that reason about the world rather than only recognize patterns within it.
I left energized about the field and the role efficient, deployable vision models will play in it.
JobScout — Kaggle AI Agents Intensive Capstone
Built JobScout for the capstone of the 5-Day AI Agents Intensive Course with Google (Concierge Agents track). Job searching is mostly mechanical — opening tabs across a dozen boards and rewriting the same cover letter — while the part that needs judgment is a small fraction of the time spent. JobScout automates everything except the last step.
An orchestrator interviews you, then dispatches specialist sub-agents for search, scoring, drafting, and review. Every board sits behind one custom MCP server (stdio, official Python SDK) that normalizes eleven sources — RemoteOK, The Muse, Remotive, Arbeitnow, Greenhouse, Lever, Ashby, JSearch, Adzuna, USAJOBS, and an opt-in LinkedIn path — into a single Job schema and dedupes them, so the agent integrates with one tool surface no matter how many boards exist behind it. Search is outcome-driven: it keeps deepening, rotating query keywords and paging further, until enough genuinely new jobs survive the filters.
Each surviving job gets an archetype tag and a deterministic ghost-job/scam check, then per-dimension scoring against a PII-masked resume with written rationale. Strong matches get a tailored cover letter — in your own voice, if you upload writing samples — critiqued by a second fresh-context reviewer agent, plus an ATS-safe CV PDF rebuilt from your real resume with nothing invented.
Security was the design constraint, not a layer on top. There is no code path that submits an application, which mitigates the Confused Deputy problem directly: the agent wields your authority to search and draft, never your authority to apply. Name, email, phone, and address are masked before any LLM call and reinjected only at the final local render. Dealbreakers and thresholds stay in deterministic Python so an untrusted posting cannot prompt-inject past them, job descriptions are wrapped and treated as hostile input, the MCP server is GET-only, and every tool and model call lands in an audit log. The eval suite includes a prompt-injection case that passes only if the injection fails.
The build process is part of the point: the spec was written first, and an AI coding agent implemented against it. Structure scales; vibes don't.
PhD Research Day — Poster Presentation
Presented research on Digital Twin systems for Mild Cognitive Impairment (MCI) at the University of Denver PhD Research Day.
Summit Hack 26 — 2nd Place
Built an e-waste sustainability assistant using computer vision and RAG to identify electronics and return carbon footprint, recoverable value, and safe disposal guidance.
Joined PhD Program at University of Denver(Dr. Mohammad Mahoor ↗)
Began doctoral studies in Computer Science under Dr. Mohammad Mahoor, focusing on deep learning, computer vision, and multimodal foundation models.
Papers I'm Reading
Paper reading log coming soon — check back shortly.
CVPR 2026 — Key Takeaways
Three themes defined the conference for me, from the deployment story of today to the research frontier of tomorrow.
Edge AI and efficient inference are the dominant deployment story
The biggest shift was from "can we build it" to "can we run it on real hardware." Quantization, model compression, and hardware-aware optimization (TensorRT, SNPE on Snapdragon and Jetson) were everywhere. Apple led the on-device push — deploying LLMs on iPhone via llama.cpp and Metal, and tiny VLMs on Snapdragon, with Oncel Tuzel presenting on the frontiers of on-device AI and their AToken unified vision tokenizer. The pattern is quantization plus memorization-style efficiency: cache and reuse computation, keep the active footprint small.
Apple · Oncel Tuzel · TensorRT · SNPE · llama.cpp + Metal · AToken
Selected papers
Diffusion models have become a general-purpose modeling tool
Diffusion is no longer just for image generation. Three dedicated tutorials, plus EDGE workshop talks from Stefano Ermon (Stanford) on diffusion inference acceleration and Han Cai (NVIDIA) on post-training acceleration. The conceptual expansion is the real story — diffusion now powers discriminative classification, 3D pose via flow matching, dynamic scene reconstruction, and world models. Kaiming He's "Back to Basics" even challenged the core noise-prediction objective, arguing models should predict the clean image directly.
Stefano Ermon · Han Cai (NVIDIA) · Kaiming He (JiT / Back to Basics)
Selected papers
World models and physics-aware models are the future of CV
If edge efficiency and diffusion were the present, world models were the consensus future — dozens of workshops on world models, video world models, and world-model-driven VLA systems. Jack Parker-Holder (DeepMind) presented Genie 3, toward open-ended world creation. The key sub-trend is physics-awareness: pixel prediction alone is insufficient, models need physical and causal structure, reflected in CVPR papers like PHANTOM (physics-infused video generation). Thomas Serre's keynote reinforced it from neuroscience — scaling alone won't close the gap with biological vision; we need "neural laws," not just scaling laws.
Genie 3 (Parker-Holder / DeepMind) · PHANTOM · Serre keynote