telecomhacker

Soooo I've been working on a recommendation system for a top 50 shopping app in the app store. Was looking for ways to brainstorm ideas that could be used for my use-case, so I took 11 papers from 2025 RecSys and fed that into LLM's to get a summarized output that also looks for content relative to the site (tunable in the prompt.json).

Wanted to share in case other folks would like to try. Some ideas was to feed best practice coding books in there and creating an md file out of that. Anyways hope it helps you folks out there!

What an insane ride it's been building this thing.... After coming back from a 2 week trip to Japan to visit my old study abroad friends and professor this past May, I decided to build https://nipponhomes.com then https://nipponhomes.com/analytics . Out of transparency, I'm losing like $2k/year in raw cloud costs (Vercel, Supabase, Zyte, OpenRouter, AWS, etc.) but I am so humbled in how much I've learned just by building this. Tbh, I realize that I am so fortunate to have the time and money to experiment with something like this, as I know many people out there are really struggling with the current market (heck I am getting rejected left and right too by my dream companies). Documenting here some of my learnings to share:

1. I learned how to create a lightweight, custom multi-modal recommendation system; I also ended up getting 2nd place for in Liquid AI's hackathon with this. (https://github.com/angelotc/lfm2-vl-embeddings) . Turns out you just need an MLP or attention layer to fuse two sets of dense embeddings.

2. Queues + async workers are a must for processing things at scale (listings in my case). Kinda go into it more in this video: https://youtu.be/qXOk7_3vZgQ?si=Mk1l3dYhzdQuvFe3&t=360

3. You need proxies (Zyte, BrightData, OxyLabs, etc.) to scrape at scale if you don't want to build your own proxy rotation system.

4. Wasn't getting sign ups until I added this feature where after a person views 3 listings, they have to sign up. That like 10x'd my signups (#growthhackingiguess)

Ps. I kinda built this out of depression tbh lol as I got rejected to Meta for the 2nd year in a row and Open AI for the 3rd time. The site currently has 8k monthly users, which is super cool, but tbh I don't know if I want to keep working on it anymore as I'm not really learning anymore, and just adding shit here and there. I know the site isn't perfect yet, and I'm getting some interests from major banks, japan real estate consultants ( the folks that help you buy the houses), and competitors (they want the data) in case you folks were interested on who is reaching out.

It’s purely remote and super chill. Not everyone wants to work on ads/compete with Indians/Chinese. I’d rather make 200k helping clinicians be more efficient using ML than $300k+ optimizing two tower models to increase the CTR on ads.

I primarily work on clinical data, and from that side, the technology stack—MUMPS included—has its quirks but generally gets the job done. The real dysfunction in U.S. healthcare isn’t the software or the language itself, but the system it’s built to serve. The core issues lie in the incentives around revenue cycle management and the structure of the insurance industry. Blaming MUMPS is like blaming COBOL for bank fees—it’s the system, not just the syntax.

Part-time MUMPS programmer here for a health system in NYC. I still love writing in it. The rates are way better than other eco-systems (e.g. Python, Java, blah blah) , probably because the eco-system isn't diluted with low-wage workers from India/China. This is because 95%+ of Epic/Ex-Epic employees are American. I would even argue it is the patriotic language of choice due to that reason.

Expected pay of 85-120/hour, which pays way more than my full-time job. It's a fun language to write in, and the adrenaline rush you get when you get a triple index loop working is awesome.

Also random fact - according to Epic HR , the average college GPA of Epic employees was 3.5, which is probably the perfect formula in hiring loyal corporate servants. I always thought it was weird that I had to apply with my transcripts and resume.

You guys should try and get acquired by Cognigy.

The reality is - businesses most of the time already know their most asked questions from their customers (based off of feedback from call center agents) when they're asking a voice bot. E.g. - What is the status of my order? How much do I have left on my balance? Can I please pay my balance off?

99% of the time, we can just build a simple intent flow off of dialogflow pointing to the customer's API endpoints that will return that data. No where here do we need an LLM / RAG since their endpoint already points to that answer. Hope that makes sense!

I work in the telecom space. I don't think this paradigm will get adopted in the near future. Customers are already building voice bots on top of Google Dialogflow e.g. Cognigy. Cognigy does have LLM capabilities, but it is not widely adopted. I think voice bots will still have to be manually configured for some time.