Alessandro Ciffo

A few months ago I launched @polymarket_data. Started as a fun side project, hit $10K MRR one month from launch, now growing steadily. Since I want to try this build in public thing, here's the full story so far. Last summer I started playing around with algorithmic trading strategies on Polymarket. Automated trading is a research-heavy process: you have a hypothesis for a strategy, you backtest it on past data to analyze its performance and then if you're confident enough you start using it to trade live. Therefore historical data is key. So I started looking for paid providers for historical Polymarket data, but to my surprise I could find none at the time. Polymarket has an official API, but it doesn't provide historical data for resolved markets. It also doesn't provide historical order books data, which was crucial for the strategies I wanted to test. Since data is what I do, I rolled up my sleeves and started collecting and processing the data myself. This seemed promising also from a business perspective, since I found a combination of something that I needed and would have happily paid for, but that didn't exist yet. (On top of that at the time I was looking for an excuse to use @ClickHouseDB on a larger project and this looked like the perfect use case. I'm a huge Clickhouse fanboy now). So I set up the database, data pipelines and monitoring stack. After a short period of testing, everything was running smoothly and the database size was growing fast. In the meantime I asked my co-founder @joaoromo_ if he wanted to give this a shot. From that moment we spent most of our nights and weekends working on this together. A few months later, in January 2026, the size of the database had reached >10TB. So we thought this was starting to get interesting enough to potentially sell access to the data. So we built a very simple landing page. The page explained what data we provided, what pain points we solved and it had a form through which customers could give more info about themselves and request access to the data. The goal at this stage was just validating the idea. Starting from the very next day people started reaching out asking to buy data. We closed multiple sales that same week, of which some worth a few thousands bucks. The idea had been validated. At that point we still had no API, but just provided data through exports on S3. This was a major bottleneck, as each export required negotiating with the customer and then doing custom work to prepare the data and serve it. So we started building the API to serve the data at scale, while still providing bulk exports for bigger customers. At the end of February we launched the API. We reached $10K MRR one month after launch and now we are growing steadily. We did all this as a 2-person team. I believe one of our main strengths was that we used AI agents since day one to automate as much as possible of business operations (things like support, sales, marketing, SEO and so on). In the last 3 months we kept working on PolymarketData every minute of our free time. Trying marketing strategies, new sales channels, maintaining the API and a lot more. There is a lot of stuff I left out in this first post, so if there’s anything you’re curious about let me know

@lennysan @danshipper However doesn't the number of humans required decrease as more things get automated?

.@danshipper: "Automation is a lie. Every time you automate something, you need a human on top of it, making sure that it continues working."

One of the things that gave us the largest productivity boost at @polymarket_data was setting up an agentic workflow to connect customer support and engineering using @FeaturebaseHQ and @linear. This makes us save a ton of time and allows us to ship faster. Here’s how we built it. Our main tool for support is Featurebase. This gives us both the support dashboard, but also a chat interface that lives both on our website and on our API’s dashboard, and that customers can use to ask questions. The tickets we receive break down more or less like this: 80% are questions about what we offer or how to query the API for specific data. These are generally easy to answer and an LLM with enough context can answer accurately. 10% are requests for custom data exports. That is, people asking us if we could provide some custom-made datasets containing data that isn’t necessarily included in our API. The remaining 10% is divided between feature requests and bug reports. This is the most interesting part. The 80% of easy questions is quite easy to automate and has been so for some years now. The agent has access to both Featurebase and our knowledge base and replies to support tickets. There is a skill to orchestrate and automate this. The only tricky part here is making sure that the agent only replies to questions for which it has enough context. But at this point in time models are smart enough to do it quite well. The 10% of custom export requests can’t be automated by definition. They're one-off requests and if they were recurring, they'd already be in the agent's context and it would just answer. So we have to handle these ourselves. Where agents really shine though is the remaining 10% of bug reports and feature requests. Here the main units are @linear issues, which can be bugs, features and improvements live and are picked up for development by some other agent. The way we automated this part is the following: If the support agent gets a bug report → It checks bugs on Linear. If there is already an issue for that, it upgrades the priority or adds more context if needed For feature requests → We have a page in our knowledge base that collects all requests. The reason why we don’t add them directly to Linear is that we believe a single request shouldn’t become an issue right away, because maybe it represents a single user’s preferences or it’s something we don’t want or can’t build for whatever reason. So this page acts as a staging area collecting feature requests and a numerical counter of how many times it’s been requested. If the counter goes above a certain threshold (we are experimenting with this), then it gets added to Linear. This is work in progress, so we still watch it closely and and keep iterating. But it’s crazy the amount of time and efficiency boost it’s made us gain. Are you integrating AI agents to automate your company’s operations? If yes what are the workflows that have given the biggest productivity boost? P.S. I know, the diagram is ugly af, but I was getting bad results with AI-generated pics so had to draw it myself

@GiulioVaccari @polymarket_data I'm in bro there's still so much inefficiency to exploit on Polymarket. We just need a few millions to get started

@aleciffo @polymarket_data We should start an hedge fund

Was curious whether Polymarket misprices Formula 1 pole sitters right after qualifying. So I backtested a simple strategy on historical data from @polymarket_data. Premise: from 2021-2025 the pole sitter won the race 54.4% of the time. So if Polymarket prices the pole driver below 0.544 in the race-winner market right at the end of qualifying, that's a mispricing and hence buy. Tested it on 13 races from Aug 2025 (when our historical order books database starts) to May 2026: - 4 trades fired, 3 won - +70% ROI on capital deployed - Best one: Antonelli at 0.34 in China 2026. +194% on that trade alone Another interesting finding is that in those 13 races, 11 pole sitters actually won the race (84.6% vs the 54.4% historical baseline). Polymarket seems to be systematically underpricing the front row. Live test for tonight: Russell took pole at the @F1 Canadian GP last night. Polymarket priced him at 0.48 right after Q3, and he's at 0.42 this morning (well below the probability estimated from historical data), so the strategy fires. Let's see how it plays out.

@alxfazio Or you can't see shit but refuse to wear glasses like me

@agazdecki The lows are what make the highs so good. It's a loop of unpredictable rewards that makes it addictive

The highs are addicting and the lows are extremely rough but one big win somehow convinces you all the stress, uncertainty and chaos is worth it. If you're a founder reading this, keep going.

@localghost To be fair sometimes in the multi-option interface the content/quality of the options is so similar that an "indifferent" button or smth like that would be more useful

Watched a normal person use ChatGPT and I am pretty confident everybody is picking from the RLHF options at random

Hey @nikitabier, could we add a feature to play videos at 2x speed when the user presses and holds on the player?

Just remembered that the @enhanced_games start next week. Really curious to watch them. I'm bit surprised that @Polymarket doesn't have any markets for those events tho

@GiulioVaccari @Dipper_pol Cause they're ashamed

@aleciffo @Dipper_pol Why are they hiding then

The secret of Hedge Funds is revealed in a 17 page PDF Stanford released the complete Hidden Markov Model framework that quants at firms like Jane Street & Caissa Capital are known to use & released it for free. Watch the post below before someone takes it down

@shiri_shh I'd take it for granted that whatever data OpenAI (or any other big tech) has, the NSA has it too

No government on earth has the amount of data OpenAI has right now not the NSA not china not even Google In Jan, they launched ChatGPT Health. user started uploading medical records, lab results, everything. Yesterday, OpenAI launched finance. bank accounts, every transaction, your net worth. and people were ALREADY dumping their secrets, insecurities, 3am thoughts into it they're building the most complete profile of human beings that has ever existed people are GRATEFUL for it btw 😭

@shiri_shh Is the idea behind buying Bitcoin miners that they can be converted to data centers? If so how easy is it do in practice?

> be Leopold Aschbrenner > german kid, Columbia valedictorian at 19 > get hired by OpenAI's superalignment team at 21 > write a memo about security being a joke > get fired in april 2024 for "leaking" > lol ok > two months later drop a 165 page essay called situational awareness > predict AGI by 2027 > essay goes nuclear in sf > patrick collison, nat friedman, daniel gross slide into the dms > start a hedge fund in september > name it after your own essay because why not > everyone else is buying nvidia, microsoft, openai > you buy power plants, fiber optics, bitcoin miners, ssd makers > $254m → $5.5b in 12 months > you are 24 > mfw the company that fired me made me a billionaire

@hubermanlab What are the best exercises to keep lower back and knees strong? About neck training, I heard it increases risk of sleep apnea. Is that true?

The big stuff: e.g., regularly doing compound (multi joint) exercises with progressive overload, moderate pace and high intensity interval training etc are made possible for decades longer if you also do the small stuff: keep your lower back, knees, rotator cuff and neck strong.

@GiulioVaccari So cool haha

We went to a robotics hackathon and wanted to do something with cameras. We derailed and made a thing to take selfies with London TFL cameras. No prize but we got some great shots of elephant & castle: londonselfiecam-virid.vercel.app

@gdb It's very good until it applies changes to your code but the revert button doesn't work (happened 3 times to me in the last few days)

codex app is very good