This program offers $500 grants per quarter to individuals pursuing advances in the field of AI models. Ideas that can become financially self-sustaining will be preferred.
There is also the opportunity to apply for a $10,000 investment.
Who can apply?
Individuals of any age and location that have an active GitHub/GitLab account.
What initiatives are eligible?
Any initiative that progresses the development and use of AI models is eligible. This might include:
- Training
- Fine-tuning
- Inferencing
- Other tools for deploying AI models
What is the grants and investment process?
- Fill out this application form. Expect to hear back within 30 days of applying.
- Selected applicants will be emailed for a 15-minute interview on a rolling basis throughout the quarter.
- Grants/Investments will be offered to successful interview candidates on a rolling basis.
How are grants paid?
Grant recipients will be required to provide an invoice to Trelis LTD and will be paid in the following preference order.
- Bank transfer to any of the following countries.
- The grantee providing a Paypal or Stripe invoice.
- As a last resort, payments can be made in the form of Runpod credits (for GPU rental).
Are there any strings attached to grants?
On completion of their grant, grantees are required to submit:
- A Title/Headline for what they accomplished.
- A 280 character tweet-style summary.
- A 500 word summary.
With the grantee’s permission, Trelis will announce the grant recipient’s name, project name and tweet-style summary.
Trelis does not claim any rights to the work being done by the grantee for the grant.
When are grants announced?
Grants are announced at the end of the quarter following completion. For example, if a grant is completed on July 15th, it will be announced by October 1st along with all grants completed by the end of that quarter?
How are investments made and paid?
- To be apply for an investment, there must be a corporate entity in place already or within the next 30 days. Delaware C Corps are preferred.
- Typically investments will be made in the form of a SAFE.
Historical Grants Completed
-2Q2026-
Eoghan Collins. For Project Tarski – a neuromorphic compute-in-memory chip using variable resistors and capacitors for analog accumulation, targeting power-efficient edge AI.
Igor. For Exopoiesis – ML-accelerated screening of iron-sulfide minerals as proton barriers for early protocell membranes (preprint).
-1Q2026-
Alejandro Alfaro. For Indigo – an AI-native design canvas that helps fashion designers generate factory-ready specs automatically.
-4Q2025-
Omid Rohanian & Mohammadmahdi Nouriborji (NLPIE Research). For ModernALBERT: a compact, modernised ALBERT-style encoder (note: includes recursion on intermediate transformer layers, very interesting!). Project Link: NLPIE. Personal Links: X (Omid), GitHub.
Maksim Liashch. For Whisper AI – transcription bots with proprietary speech-to-text models for Telegram and WhatsApp. Project Link. WhatsApp Bot. Personal Links: LinkedIn.
-2Q2025-
Dima Yanovsky. Accelerating Robotics Imitation Learning via Simulation and AR Teleoperation. Project Link. LinkedIn. Github.
-1Q2025-
Martin Disley. For the development of a neural pruning-based method of concept suppression in transformers. Using a hybrid approach of concept neuron saliency analysis and pruning, and vocabulary neuron pruning, we demonstrate suppression of knowledge of the colour “red” in a pruned Llama-3.2-1B-Instruct model. Personal Links: X, GitHub.
Nilan. The goal of the project is to train a video tokenizer capable of compressing 256p, 32-frame videos to 128 tokens with a codebook size of 4096. The quality target is ~30 PSNR and ~0.85 SSIM. The grant will go towards compute costs for model training and dataset processing. Personal Links: GitHub
Madan Khadka. For our Video-to-VFX system, we utilize a fine-tuned CogVideo model to transform videos into basic VFX. Recently, after speaking with a new customer, they suggested adding a storyboard feature to simplify short video and movie creation while also enhancing video quality. This idea inspired me to explore its potential, and I’m excited about the possibilities it offers. Personal Links: X, GitHub
Tiberiu Toca. Building Gaudi, an agentic CAD IDE for coordinating end-to-end AEC project execution. Personal Links: X, GitHub, LinkedIn.
Ali Asaria. For Transformer Lab: adding additional functionality to support preference tuning including RLFH, DPO, ORPO, and SIMPO. Project Link: Transformer Lab. Personal Links: X, LinkedIn, GitHub
Aarav Sharma. Brain-Controlled Exoskeleton Arm for Individuals with Paralysis. Personal Links: GitHub, LinkedIn.
-4Q2024-
Roddiyyat Taiwo (X, LinkedIn, GitHub): For OpenHealth – an AI-powered preventive healthcare platform based in Africa.
Ryan Lucas (LinkedIn. Github.): For a second-order method for neural network pruning.
John Denny (Personal Site, LinkedIn, X.): Building a system to pre-emptively detect and monitor scam-like domains that imitate brands and governmental bodies, then perform automated takedowns. Project link on Github.
-3Q2024-
Fintan Parsons (X. LinkedIn. GitHub.): Solving a problem in optics and 3D printing by using AI to determine the aperture and light sources needed to create specific interference patterns.
Sadra (Sam) Sabouri (X, LinkedIn, GitHub.): For designing an accessible API for the PyCM library, making it easier for users, especially those who are not familiar with Python, to utilize the library in the evaluation phase of machine learning models. Project Link: https://github.com/sepandhaghighi/pycm
Nathan Brown (X. LinkedIn. GitHub.): Building the first open LLM capable of understanding Setswana, an underrepresented African language in AI. This includes model weights, a web text dataset, and a synthetic dataset.
AmirHosein (X. LinkedIn. GitHub.): Adding “ML Streaming” feature to PyMilo. “ML Streaming” simplifies model deployment and remote usage. This feature includes a wrapper class to abstract backend communications, making it user-friendly.
-2Q2024-
Stephanie Wan (GitHub.): For the creation of an author verification tool through training a BERT type model and, separately, through a feature-based approach. Try out her web app here.