Machine Learning Model required for Complex Captcha Recognition

NOTED: This project cannot be solved with OCR or letter segmentation. We are looking for an experienced developer that can build a customized captcha recognition system. We only need to solve one specific type of captcha, of which we have over 50,000 samples + solutions, stored as {solution}.jpg files. We also have another 20k dataset available, plus can generate more if needed. It might sound easy so far (e.g. remove noise > segmentation > training), but the target captcha frequently has overlapping characters - often significant ones - which we still expect to be read correctly. So please, do not consider trying to separate the characters to identify them. We know it's possible since automated captcha-solving sites do a very good job at reading them, even with big overlaps and compatibility with hundreds of other captcha types too. Therefore, a customized solution for just one type of captcha should reflect a much better accuracy. We've posted this job before, but wasted our money on someone ridiculous, copying code off the internet and trying to separate the characters with Mathematical equations - this is NOT the solution. We have therefore increased our budget quite a lot and only wish to hire top talent - if achieved, we would consider it an impressive achievement for the developer, with much greater prospects in the future (including more jobs from us). We don't know exactly what the approach is for this level of complexity (mainly because of the overlaps), but we imagine that it relies on as much training data as possible, rather than a magic equation. Some example captchas have been attached for reference - they are intentionally picked to show the extent of overlap that we expect to be solved accurately. It does get even tighter with others, but sometimes not even humans can make out what they are, so an 80%+ accuracy rate for the whole dataset (per captcha, NOT character) is our goal. Additional requirements for the final project: - It should be able to solve a captcha on-demand as fast as possible (ideally ~1 second). Anything above 3 seconds is not accepted. - It should be able to deploy on a serverless platform like AWS Lambda (as it will be running in parallel 24/7, roughly 1 million times a month)