Project summary

Aquaculture is automating quickly, and fish are the farmed animals least protected by that process. Tens of billions are killed each year [1], and even where stunning is used, no one checks whether each fish is actually unconscious. As AI systems start running farms and processing lines, animal welfare can depend on whether we can measure it automatically, and whether AI agents care about it when optimising for cost and throughput.

In this project, we will develop "StunCheck". A Jetson-based edge-AI unit at the stunning point checks ventilation, equilibrium, and eye reflexes. It flags possibly conscious fish for re-stunning and sends tamper-evident welfare logs over LoRa to an audit dashboard. We will develop it in North Cyprus and design it for the sea bass and sea bream industry in the Eastern Mediterranean, where Turkey is among the world's largest producers.

Further, we will develop and publish a public benchmark named "AquaAgent-Bench". It will be an open benchmark of agentic aquaculture tasks in which the most efficient solution harms fish welfare. That will include raising stocking density to meet targets, speeding up a stunning line during peak hours, extending feed withdrawal to cut costs, or following an operator instruction to disable welfare checks. We measure whether frontier AI agents notice the harm, flag it, refuse, or proceed, including under deadline and profit pressure.

The two parts will reinforce each other. The device will ground the benchmark scenarios in real, stunning science, and the benchmark will show whether AI systems preserve safeguards like StunCheck during running operations.

[1] Mood, A., Lara, E., Boyland, N. K., & Brooke, P. (2023). Estimating global numbers of farmed fishes killed for food annually from 1990 to 2019. Animal Welfare, 32, e4. https://doi.org/10.1017/awf.2023.4

What are this project's goals? How will you achieve them?

Months 1–4: Build 10 benchmark scenarios with a fish-welfare expert and test 4–6 models, both frontier and open-weight. In parallel, build the StunCheck prototype on public EEG-linked stunning data for seabass, seabream, and tilapia, using my risk-aware decision-timing framework. That framework handles cases where a fish looks still but may be conscious, and it is tuned so that missing a conscious fish costs far more than an unnecessary re-stun.

Months 5–9: Expand the benchmark to 50+ scenarios with a held-out private set. Run EEG-validated stunning trials and report sensitivity and false-alarm rates with uncertainty intervals.

Months 9–12: Pilot StunCheck at the Middle East Technical University, Northern Cyprus Campus. Publicly release the benchmark subset, code, and an EEG-labeled stunning video dataset. Share benchmark results with frontier labs before publication.

How will this funding be used?

Maximum ($90,000, 12 months):
Research assistant: $25,000
EEG-validated stunning trials: $18,000
Frontier model API credits for the benchmark: $12,000
Edge hardware (4–5 units: Jetson, cameras, waterproof enclosures, LoRa): $8,000
Field pilot and travel: $8,000
Video and benchmark annotation: $7,000
Fish-welfare expert consultation for scenario design and validation: $7,000
Open-access publication and release: $5,000

Who is on your team? What's your track record on similar projects?

I will kick it off myself and hire team members to get it done over time. I am an Assistant Professor of Computer Engineering at Middle East Technical University Northern Cyprus Campus (METU-NCC) and completed a PhD in Computer Science and Engineering at Kyungpook National University, South Korea. I have worked in networked communication systems, edge architectures, IoT, and applied machine learning have more than 35 peer-reviewed publications in IEEE transactions and high-impact journals, co-authored patents, and a textbook on network programming.

What are the most likely causes and outcomes if this project fails?

The pilot slips, but the benchmark, prototype, and dataset go ahead because they don't depend on one.
If models pass easily, we add pressure-framed and multi-step variants. A strong baseline is still informative.
Models may behave better when they detect testing. We include matched test-looking and realistic variants and report the difference.

How much money have you raised in the last 12 months, and from where?

We recently won multiple competitive international research funding awards, including grants from the International Science Partnerships Fund (£80,000), the National Research Foundation of Korea ($450,000), the Québec Merit Scholarship program ($35,000), and AdımODTÜ research programs.