This is not a traditional internship. You will not be paid but also you will not be given tightly-scoped tasks. You will be given a problem and you will be expected to produce a demonstrable artifact (code, paper, dataset, or running service) that survives an honest code review.
Eligibility
The program is open only to applicants who meet all of the following:- US work authorization. You must be authorized to work in the United States for the duration of the program (US citizen, permanent resident, or an appropriate visa status that does not require Aegean AI to sponsor).
- Graduate-level enrollment. You must be a currently-enrolled graduate student (Ideally PhD but we also consider MS) at NJIT or NYU.
- Minimum qualifications You must have completed at least a graduate level AI course and have secured an A grade. Ideally you are very well versed with NLP, CV or Robotics projects.
- Minimum compute You need to have access to 30, 40 or 50 series consumer-grade GPU with at least 12GB of VRAM. You also need to have a subscription to Claude Code or Codex (we currently prefer Claude Code). Unfortunately, Google Colab Pro does not count.
Domains of interest
Projects are drawn from Aegean AI’s portfolio. In 2026 the priority domains are:Reasoning & RL post-training
GRPO, DPO, reward modelling, async RL loops, reasoning evals.
Vision-Language & Vision-Action Models
CLIP / LLaVA / OpenVLA-style systems, multimodal reasoning, grounded tool use.
Robotics Systems
Perception, planning, and control integrated with LLM-based agents; world models (eg. JEPA family).
Physical AI
Edge deployment, sim-to-real, embodied evaluation harnesses.
Program rules
These rules exist because the program uses coding agents on real code under our licence and liability.- Human-in-the-loop is mandatory. Every pull request the agent opens is reviewed, tested, and merged by you under your own name. Unreviewed agent output is never pushed to a shared branch.
- Cite the agent. Every commit message or PR description includes a
Co-Authored-By:line naming the specific coding agent (model + version) used. Transparency is non-negotiable. - No scraping, no un-licensed data. All training or fine-tuning data comes from licenced or first-party sources. If you’re unsure, ask before running the job.
- Open-source by default. Unless the project lead declares a carve-out in advance, all code you and the agent produce is released under the project’s existing open-source licence. This includes notebooks, eval harnesses, and prompts.
- Evaluation is a deliverable, not an afterthought. Every project ships with at least one evaluation artifact (benchmark script, held-out set, reproducible notebook) that someone else can run to confirm the claimed result.
- Weekly demos. You present progress every week: what the agent did, what you corrected, what went wrong, what you learned.
- Commitment: During the summer we will help you achieve the project goals and you commit to work at least 8h/workday on this project. If a publication is warranted, for a conference, you agree to publish the work with you as the first author but with a member of technical staff of Aegean AI Inc as the coauthor. If the end result includes a demo, you agree that will also be published as part of the products/tech-demonstrators/ page.