we are working with this consumer electronics company debug their zephyr project right now, would be worth trying out. if it doesn't work, email [email protected] we'll help you out
we are talking to a some of the larger vendors about getting first-party support for documentation. once we scale up a bit I think it could make sense to have multiple fine tuned models based on the docs they provide (or we crawl the web for public ones, at minimum)
there are some other companies trying to solve for PCB design. long term we want to build end-to-end AI automation for those cross-domain projects. this was a good place to start based on our personal/work experience
need to update the policy, this was for the original version of the app, which used a deterministic algorithm to generate driver code, which we stored on cloud so you could download the files. thanks for pointing this out.
strict grounding protocol + planning phase, mostly by prompting and forcing attention through citations. it tends to think longer than other coding agents but the results are usually better. let me know what you think.
great question. we've found most code/docs are offline in the embedded space (for good reason) so our approach going forward is going to be more FDE/on-prem for enterprise users. they asked for self-hosted, BYOK, local indexing, etc. and I think this is something that can differentiate us
for consumer users we have a zero-retention policy with the model providers, and we use repo mapper to index your code locally, but as you pointed out these APIs are a black box so no guarantees
currently we scaffold frontier models. the product is basically a context layer with custom tools that enable hardware interaction. we've tossed around the idea of pre-training/fine tuning but new models are being released so fast it doesn't make sense to build anything other than a wrapper