#business #ai #llm Created at 190223 # [Anonymous feedback](https://www.admonymous.co/louis030195) # [[Epistemic status]] #shower-thought #human-in-the-loop Last modified date: 190223 Commit: 0 # Related # TODO > [!TODO] TODO # How I built Jasper ai for B2B As co-founder and [[CTO]] of Inkdrop AG, I built a production description generator for e-commerce resellers. We were part of Techstars Berlin 22 batch. The product consisted in a web app allowing freelancers paid by us to work in copilot mode with [[Artificial intelligence|AI]]/[[Large language model|LLM]] to write a large quantity of product descriptions asked by our customers. Customers could use our web app to track, export, review, accept and reject products (we would use this feedback to improve our models). After a few weeks into Techstars batch we decided to drastically pivot and say farewall to our $40k ARR customer contract, quickly after I left Inkdrop for new adventures. > [!TODO] [[ChatGPT]] summary > The author describes their experience as co-founder and CTO of Inkdrop AG, where they built a web app that generated product descriptions for e-commerce resellers using AI/LLM and human co-pilots. The app included features such as online product retrieval, integration with Google Search Console, IAM, roles, and permissions, a model zoo, real-time feedback during human-in-the-loop, templates, SEO optimization, and a roadmap to full automation. The author also mentions using a distributed crawler for a million-scale e-commerce website and collecting human interactions to reduce LLM hallucinations. ## Online product retrieval We had a product called "Krafla" which would search google to find similar product, and parse the top 5 results, format and generate automatically missing fields. There is a big problem that we tried to tackle, that is being able to parse any web pages using a LLM. Multiple things tried: - Parsing chunks of HTML into properties (because LLM input is limited) - Screenshot -> OCR -> parse properties (because the vision part is what humans see, the most important information we needed) ## Extra messy information ### Integration with Google Search Console We were letting the customer see how our product descriptions influenced his revenue through charts embedded in our product. We also explored using SEO data to train models to predict a product description SEO beforehand. ![[Pasted image 20230219184830.png]] ### IAM, roles, permissions In the product we could create users with roles. We could create roles and assign some permissions, these permissions would dynamically adapt the UI. ![[Pasted image 20230219184608.png]] ### Model zoo ![[Pasted image 20230219184456.png]] ### Inkdrop product workflow ![[Pasted image 20230219184356.png]] ` ### Realtime feedback during human-in-the-loop ![[Pasted image 20230219184302.png]] ## [[Product discovery]] framework in Inkdrop ![[Pasted image 20230219184231.png]] ### Templates Inkdrop had a feature to let customers create template for streaming their products to our system in the most frictionless fashion and get their products written in their favorite way (for example some customers wanted technical details, some not), optimised for their preferred metrics (rewards humans according to customer goals ...) and seamlessly integrated into their system ![[Pasted image 20230219183930.png]] ### SEO optimized [[Large language model|LLM]] ![[Pasted image 20230219183816.png]] ![[Pasted image 20230219183856.png]] ### Roadmap to full automation ![[Pasted image 20230219183717.png]] ### Scrapping into Inkdrop product ![[Pasted image 20230219180437.png]] ### Collecting human interactions to reduce [[Large language model|LLM]] [[Hallucination]]s ![[Pasted image 20230219180313.png]] ### Highlight category ![[Pasted image 20230219180148.png]] ### Distributed crawler for million scale e-commerce A similar architecture has been used to crawl all products of the biggest German e-commerce websites. I played around with this data on my weekends (it was not really on our priority list), for example to train [[Self supervised learning]] sentence embeddings models like https://www.sbert.net/examples/domain_adaptation/README.html#gpl-generative-pseudo-labeling ![[Pasted image 20230219175841.png]] ### Retrieval augmented generation ![[Pasted image 20230219175548.png]] ### Word repetition user feedback ![[Pasted image 20230219175511.png]] ### Ban words user feedback ![[Pasted image 20230219175444.png]]