After doing a bunch of projects for everything but AI Voice Agents, I finally gave into the hype. Some weeks ago, I showcased the final demo for a custom-coded AI Voice System made for one of the biggest IoT Companies in Spain, and they seemed to love it! Now I know: the ticket is not that high, specially for such a big company. But, as stated, this is the first time I dive away from RAG systems & other solutions, and go into AI Voice Agents instead. And I didn’t want to disappoint! (and no, sadly I couldn't make it a recurring subscription). Managed to get this deal thanks to a professor from a University in Spain who leads research projects in collaboration with this big IoT company, and this should be the first (and cheapest) phase of the project. We're now heading into automated document generations using OpenAI + Google Docs (for URDs, SRDs, etc.) Fully aware this is not the most impressive deal, but we're building things up slowly, there's more to come. ----------------------------------------------------------------------- Project info: 📌 What is it? Custom-built AI voice assistant that initiates interactive, outbound calls at the time desired by the user. It calls the clients from the IoT company who want to develop software projects with them and gathers all the necessary information to get started with the project and create the necessary documents (URDs, SRDs…). ⭐️ What does it solve? What they have to do without the voice agent is: fly out an employee to wherever the customer is located, and speak with them directly to gather all information they require to get started developing the software. An incredibly expensive and time-consuming task for the business. Now they can just send a form to the customer, they fill in their information + desired time for the call (instantly or at a set date and time). Then the LLM will go through a series of questions set by the company, and it will keep asking until it gathers all the necessary data from the customer.

Diving into outbound AI calling these past few days using OpenAI’s new Realtime API has been… quite a ride. Anyone who’s tried it will know that outbound resources are basically nonexistent (only seen some docs on inbound calls), so I had to dig deep and figure things out on my own. I decided to make a video sharing the full breakdown of what I learned. Not going to lie, it’s my first AI-focused video, and there’s room for improvement, but I wanted to get something out there that’s actually helpful (even if its not a masterpiece) Take a look at the vid if it sounds interesting, somehow its already at 200+ views Here’s the GitHub repo too if you want to check out the code directly: https://github.com/MarcosSan4/realtime-twilio-outbound.git If you’re diving into similar challenges or have any questions, feel free to reach out. Happy to share what I know and learn from each other!

Picture this: You’ve launched your AI solution, and it’s being put to good use. But suddenly, you’re overwhelmed, manually reviewing every new user query to ensure everything works as intended. The flood of queries makes it nearly impossible to spot repeats, and it’s eating up your valuable time. .... What if you could automatically identify new, unique queries without lifting a finger? After facing this challenge myself, I developed a Python-based system that got rid of a lot of headaches. - 📚 Automated Similarity Checking: By using cosine similarity (or similar methods), you can automatically compare new user queries against previous ones to identify duplicates or closely related questions. - 💬 Efficient Query Management: The system flags truly unique queries, allowing you to focus your attention on those. - 📈 Scalable Solution: As your user base grows, this approach scales effortlessly, handling an increasing number of queries without additional manual effort. It honestly took me a while to have the structure of the system fully laid out and working, so I can hopefully save you some time with my explanation! (the loom vid attached showcases the system a bit more in-depth) But that’s just the tip of the iceberg. You can take this even further by: - Automating Response Consistency Checks: Compare new replies to old ones for the same queries to ensure consistent responses. And integrating additional features to automate your workflows even more. fyi: there may be better ways to get around solving this issue, this is just the solution I came up with. By setting up this system, you can: - Save Time: Drastically reduce the hours spent on manual reviews. - Increase Efficiency: Quickly identify and address unique issues. - Scale Effectively: Maintain performance as your user base grows. If you’re interested in learning more about how to set this up, including detailed steps and code examples, I’ve laid it all out in a recent article.

Should have made this post when it happened, around a month ago, but hey better late than never uh. Some background to the story: for the past five years, I’ve been deeply involved in coding, developing a robust set of skills in Python and no-code tools. Initially, I spent around 3-4 years running a B2C business, where I enhanced my development skills and also gained good experience in understanding customer needs. Making the full switch to AI and developing AI-driven solutions has been an exciting journey (lots of downs with some ups), and I’ve dedicated countless hours to creating various projects, specially focusing on automated document generation & RAG Chatbots for everything from general Q&A, to closing leads, and client onboarding. And finally, at the start of September, I secured my first B2B client through warm outreach for a complex Retrieval-Augmented Generation (RAG) system I had been developing in Python for some time. → 🎯 Objective: Provide instant answers to all customer (student) queries through WhatsApp regarding a complex topic with various inter-related subjects. → ⁉️ Why: By offering a 24/7 automated solution, the client can reallocate over five employees to more significant tasks instead of having them answer the same repetitive questions repeatedly. Plus, it’s a unique selling proposition (USP) for them in a very outdated market. Plus, doing this whole process with custom code allows for the costs to be dirt cheap, thus making this whole thing very scalable. → 📌 How it works: Explained it on the video attached, hopefully it is clear enough (first time using Figma so.... not my best work that's for sure) I’m now in the process of closing three other deals, which will hopefully go through. So, if you’re looking for assistance in creating the overall architecture of AI solutions, facing challenges in development, or need to hire someone to develop a custom solution, feel free to reach out :) Either way, thank you for reading through the post. Hopefully this inspires or brings new ideas to you.

Hey everyone! A few days ago, I posted an article about a Python script I developed that helped me seamlessly transcribe and integrate 3,000+ visual slides full of crucial information into my RAG Knowledgebase—completely on autopilot. Now, I’ve given the script a full makeover to make it as easy as possible to plug-and-play! The article took off on Medium, so I knew I had to bring it here for you all to benefit: [ARTICLE LINK + SCRIPT LINK & FREE PROMPT TEMPLATES] [LinkedIn Post link in case you prefer it] This all started because I was developing a RAG chatbot for a client and... I ran into a pretty big problem. All the info he provided was in slides full of text & images! I had to figure out a way to implement these slides into the Knowledgebase (and it had to work). So I got to work and developed a script to do so on autopilot. It leverages the vision capabilities of the latest LLM models to transcribe the slides accurately—OCR just couldn’t cut it because it ignores the layout and misses out on crucial images. But that’s not all! The script is loaded with extra features: - SUMMARIZING & VECTORIZING: The script doesn’t just transcribe; it summarizes key concepts and creates vectors to ensure your Knowledgebase captures everything. These vectors are essential for data integration. - FOLDER PROCESSING: It processes every subfolder in your directory, so no image gets left behind. Perfect for managing large datasets. - SMART FILE NAMING: The script updates transcription filenames with vector counts, so you always know where you stand. - MERGING TRANSCRIPTIONS: You can merge all transcriptions into a single file—whether by folder or into one master file—keeping your data organized and accessible. - VECTOR COUNTING: Get a quick snapshot of your data volume with vector counts for each main folder—great for ensuring completeness. - VECTOR UPLOADING: Finally, it uploads all vectors to the Qdrant vector store (but you can switch to another provider with a simple code tweak).