Applied ML Research Engineer Intern — Game Portability & Runtime Control (Unpaid)

Help test new learning methods against real systems problems

VenturSeed is developing DartStream and IntelliToggle. We are looking for an Applied ML Research Engineer Intern to investigate whether learned representations and planning methods can improve game portability and runtime control.

This is an unpaid research internship. There is no promised salary, stipend, equity, or conversion to a paid position. The proposed initial research scope is an 8–12-week feasibility project; start date, weekly hours, location eligibility and final duration will be agreed before participation.

The research

For DartStream, investigate how models can help adapt PC games to different hardware: choosing useful compatibility or configuration changes, reducing manual trial and error, and evaluating results on game/hardware combinations not seen during training.

For IntelliToggle, investigate how models can propose precise, dependency-aware configuration, shutdown and recovery plans. Experiments will use bounded test environments. Production execution remains subject to deterministic validation, permissions and engineering review.

Next-Concept Modeling (NCM) is one candidate approach to evaluate alongside conventional methods and suitable open-weight models. We are not assuming NCM will outperform those alternatives. A well-supported negative result is valuable.

What you would work on

• Turn a narrowly defined research question into reproducible experiments with clear success and failure criteria.
• Curate versioned datasets from approved test cases, execution traces and measurements; prevent leakage between training and evaluation sets.
• Establish strong baselines before implementing experimental representations or model components.
• Train, debug and evaluate models in Python/PyTorch within an agreed compute budget.
• Test generalization on withheld game/hardware pairs or dependency configurations, including failure and recovery cases.
• Compare accuracy, constraint violations, latency, memory use and experiment cost at comparable resource budgets.
• Document code, configurations, results and limitations so another engineer can reproduce the findings.

The scope is a focused research prototype. You would not be expected to independently own the entire product, production safety architecture, hardware infrastructure or game-porting pipeline.

What strong fit looks like

• Hands-on Python and PyTorch experience, including training and debugging models beyond calling a hosted API.
• Evidence of reproducing a paper, building an experimental ML system or completing a substantial research project.
• Understanding of evaluation design, held-out data, ablations, fair baselines and common sources of misleading results.
• Ability to work within GPU memory and compute constraints and explain the tradeoffs you make.
• Clear written reasoning, reproducible code and willingness to report results that contradict an initial hypothesis.

Experience with transformers, learned discrete representations, vector quantization, compilers, graphics, compatibility systems or systems telemetry is useful but not required across every area. We welcome applicants with strong project evidence, including graduate students; a particular degree is not a substitute for demonstrated ability.

Proposed first project

First, select a bounded test problem and establish the dataset, evaluation protocol and baseline. Next, implement and test one experimental approach. Finally, compare results on unseen cases and produce a concise recommendation to continue, change direction or stop.

Expected outputs are a reproducible repository, documented data and evaluation splits, an experiment log, a results report and a research prototype suitable for engineering review. Compute access and any approved expenses will be agreed before work begins; applicants should not assume they must purchase hardware or cloud credits.

Please include four short answers

1. Link one relevant repository, paper reproduction or project. What did you personally implement, and what worked or failed?
2. Describe how you would evaluate a model on unseen game/hardware combinations or dependency configurations without leaking test information into training.
3. Describe a training, GPU-memory or experimental-design problem you debugged and how you verified the fix.
4. What is your availability, preferred start date, weekly time commitment and working country/time zone? Please confirm that you understand this is an unpaid opportunity and tell us any academic or participation requirements you would need clarified.

Brief, specific answers are welcome. Please do not share confidential work or prepare a new unpaid project just to apply. We will review written evidence before scheduling an introductory conversation.