Project summary

Plantcore.AI builds AI agents for industry. Industrial requirements are different: people need reliable behavior, control over what an agent can do, and a visible engineering version whenever the system changes. We built Iteron because the general-purpose harnesses we tried did not meet our deployment requirements.

Our first agent scheduled production for an apparel factory. We had to keep changing the harness to make it useful, and when a new model came out we had to optimize again. That is the problem we want to study: can the harness be optimized for a particular task without tying it to one model, while keeping changes reviewable by people?

Iteron is open source under Apache-2.0: https://github.com/Plantcore-AI/Iteron. Its three priorities are controllability and auditability, task-specific optimization, and model independence. One customer is building a cluster to host open models. They want to choose their models and know exactly which version of the surrounding agent software is running.

Goals and approach

We are requesting USD 20,000 for a three-month pilot. We are turning our deployment experience into an open simulation environment, initially using synthetic or releasable industrial-software issue tickets. The concrete starting case is repair of tickets for a PLM system used by a battery manufacturer.

Month 1: prepare fixtures and validators, group related tickets into development and held-out sets, and freeze the evaluation protocol. Month 2: compare default, tuned, and transferred harness configurations across selected open models with matched budgets. Report failures, task outcomes, latency, compute costs, and retuning effort. Month 3: test a supervised update pipeline that records a proposed change, regression evidence, a human approval or rejection, a new engineering version, and a rollback target. We will run a small reviewer experiment comparing structured evidence with a summary alone.

The public outputs will be the task environment, evaluation scripts and results, configuration versions, and the release-control prototype. These outputs are planned work, not a claim that the full pipeline already exists.

Use of funds

Proposed direct costs: USD 14,000 for engineering and research; USD 4,000 for compute and evaluation; USD 2,000 for independent validator review and pilot human-review testing. No institutional overhead is requested. The project starts after funding becomes available and lasts three months. This budget supports a bounded public research pilot, not the full cost of our company or customer delivery.

Team and evidence

I am Yitian Lyu, founder of Plantcore.AI (GitHub: Skylovingsky). Our core team has three people, all full-time. Plantcore.AI is registered in the United States and has no academic affiliation. With an FDE team in China, we report Iteron use across five factories over two months, for tasks including scheduling, production checks, and repair of errors in existing industrial software.

In two development groups of ten historical PLM tickets each, the existing enterprise workflow produced final candidate patches that passed qualified independent behavioral validation on 3/10 and 2/10 tickets; Iteron passed on 10/10 and 8/10, with the underlying model held fixed. These are development results, not held-out evaluations, customer acceptance, or production reliability. Multiple workflow components changed, so the experiment does not isolate the effect of parameter optimization.

What could fail

Tuning could overfit the fixtures; improvements could disappear on another model; validators could miss meaningful failures; and human approval could become a rubber stamp. We will use grouped held-out tasks, independently reviewed checks, bounded permissions in simulation, and regression-inducing changes in the review experiment. A negative result is still useful if we publish what failed and what it cost. We will not treat a software test pass as permission to deploy into a live factory.

Funding history

We have received no project grant funding. A USD 25,000 Sentient application is pending, and we are pursuing other grants for related open-source work. We will disclose any overlapping awards and adjust scope or budgets rather than charge the same work twice.

Why open infrastructure matters to us

I call the broader idea compute internationalism: compute is a new general-purpose productive resource, and people in different places should have sustainable local capacity and control over it. Model choice is part of that. So is control over the software that turns a model into an operational agent.

This English application was prepared with AI assistance from the founder's project description and research materials.