🗣️ Post 1: Launch HN: Promi (YC S24) – Personalize e-commerce discounts and retail offers

As posted by: pmoot  |  🔥 Points: 21

🔗 https://news.ycombinator.com/item?id=44649115

💬 Summary

Hey HN! I’m Peter from Promi (https://www.promi.ai/). We’re building a platform for ecommerce merchants to send realtime personalized discounts, optimized with AI (obviously)

Demo: https://youtu.be/BCYNCqb4fUc, Sales Video: https://www.youtube.com/watch?v=WiO1S7RBn-o

All the big tech companies send personalized discounts - Uber, DoorDash, Google, etc. In fact, I was the product lead overseeing discounts at Uber, so if you’ve gotten a promotion on Uber Rides or Eats, that was our tech. These personalization models often generate >30% more revenue vs. non-personalized discounts (cost-neutral that is), so this is a hugely impactful product.

It’s no surprise then that other merchants want to follow suit. Merchants don’t want to waste discounts on customers who would have purchased anyway. Frankly it’s not a new idea to offer a software solution to personalize discounts - plenty of other startups have entered this space with a similar product.

The biggest problem with personalizing discounts for mid-size and smaller companies has been that traditionally you rely on ‘explore’ data - data from randomly sending out discounts to a portion of the user base. But this has a lot of problems: merchants need to be large, collecting this data is expensive, training data really should be fresh (so explores should constantly be running), and if you want to try a different discount structure (e.g. BOGO instead of 20% off) you’ll need to run a new explore with the new structure.

So what does Promi do differently? We train on regular traffic and simplify the problem by just focusing on conversion rate. If we can accurately predict who is unlikely to convert and which products are unlikely to be bought, we can issue discounts without the fear of burning money on an order that would have happened anyway. One of my major takeaways from my time at Uber was that our model was mostly targeting users who had a low likelihood of converting in a given week. Quantifying how much more likely they were to convert when given a discount via explores was helpful, but not as impactful as understanding starting conversion rate.

Side note - It’s been a bit interesting launching an AI company during this hype cycle that isn’t actually using the latest and greatest LLMs. We believe more traditional machine learning still has a lot of value to add. I don’t want to say we won’t use LLMs down the road (there may be some interesting applications for developing additional features), but starting this way has worked out well for us.

There have been plenty of other challenges (as with any startup). We’ve had to figure out how to automate integrations when so many websites have custom code. We’ve had to make the model work without rich user data since the majority of website visitors aren’t logged in. A quick note in this one - we can use first party cookies to more or less track the view and transaction history, but we’ve found that one big predictor of conversion is traffic source: whether a visitor is coming from ads, email, direct traffic, google search, etc. That traffic source isn’t something as valuable at Uber (since everyone uses the app), so it’s been a bit of a tradeoff in the types of features that are most impactful.

Our model seems to be working well! We have case studies on our website showing the typical revenue and profit lift we see. We currently have tiered pricing with different quotas for the amount of revenue managed by Promi discounts.

I’d love to get thoughts from the machine learning experts in this community, though full disclosure I’m the non-technical founder. Let us know what you think!

🗣️ Post 2: Show HN: Runtime Defense Against Prompt Injection in Supabase MCP

As posted by: anand-tan  |  🔥 Points: 3

🔗 https://docs.tansive.io/blog/implementing-defense-prompt-injection-attacks-mcp/

💬 Summary

I wrote this after studying the Supabase MCP prompt injection issue. The blog shows how I built a working defense using an open-source AI agent runtime I’ve been building called Tansive ( https://github.com/tansive/tansive )

Instead of just filtering malicious prompts, I implemented role-based policies with runtime input validation that can scale across combinations of different AI tools (GitHub, Stripe, Linear, etc.).

All the code referenced in the blog is in the examples/supabase_demo folder.

I welcome your feedback — especially from folks working with AI toolchains or security.

🗣️ Post 3: Ask HN: Prompt Template Engine?

As posted by: tmaly  |  🔥 Points: 2

🔗 https://news.ycombinator.com/item?id=44651247

💬 Summary

As my AI usage has increased, I am finding I need a way to manage my prompts and also a way to handle complex multi step prompts.

What is your open source solution for this?

🗣️ Post 4: I built a tool that turns your website or content into viral TikTok slideshows

As posted by: Fr1tz1707  |  🔥 Points: 1

🔗 https://news.ycombinator.com/item?id=44656927

💬 Summary

Hi HN! I'm 21 and recently launched my first real SaaS: https://shortgen.io

It lets you:

Enter your website, product link or Promt

AI analyzes the content

Instantly turns it into a slideshow-style TikTok (titles, pacing, text, visuals — all done)

Why? I wanted to help founders and small businesses get attention on TikTok without filming or editing anything. It started as something I wished existed for myself and I built it.

Progress so far:

815 users

$80 MRR

One feature recently went viral, bringing in 4 new trials overnight

Free trial = huge boost in conversions

also i added a feature to create clips from any video like youtube by analyzing the content with Ai.

Still very early, but I’m learning a lot fast

Tech stack:

Frontend: Next.js

Backend/API: Supabase (Postgres + Auth)

Hosting: Railway

AI/LLMs: OpenAI API + other LLM's as fallback....

Email: Resend

Logging/Analytics: DataFast

Lessons: Free plans = bad activation

Free trials (with credit card) = way better intent and way fewer support issues

Everyone says “build for yourself”it’s true. The only reason I’ve pushed through the rough weeks is because I actually use my own tool

🗣️ Post 5: Show HN: SwellDB – Query AI-generated tables with SQL

As posted by: giannakouris  |  🔥 Points: 1

🔗 https://github.com/SwellDB/SwellDB

💬 Summary

I'm building a data system called SwellDB that uses LLMs to generate its tables on the fly.

Traditional databases only work over data that's already loaded and cleaned. But in the real world, data lives everywhere — in files, PDFs, web pages, APIs. To query it, we usually need custom ETL pipelines: extract, clean, transform, load. It’s slow, brittle, and different every time.

SwellDB flips that model: you define a table (schema + a description as a natural language prompt) and it generates the table just-in-time — using LLMs and your schema/prompt, on top of the connected data sources (files, databases, LLMs, web). Think: querying a DataFrame that materializes itself from raw input without you writing the ingestion logic.

It supports:

- Structured + unstructured sources: CSV, SQL, web search results (PDF to be added soon)

- Declarative table definitions in Python

- Output compatible with any SQL query engine (DuckDB, Apache DataFusion) or ingestible into any database

Repo: https://github.com/SwellDB/SwellDB

Short paper (4 pages): https://github.com/gsvic/gsvic.github.io/blob/gh-pages/paper...

Would love feedback if you get a chance to try it out, especially from folks dealing with hybrid or messy data sources.

🎯 Final Takeaways

These discussions reveal how developers think about emerging AI trends, tool usage, and practical innovation. Take inspiration from these community insights to level up your own development or prompt workflows.