I've been building a system over the past couple weeks that sends me a push notification every morning at 9am with a full pool chemistry report, 3-day weather forecast, equipment status, and AI-powered maintenance recommendations — all grounded in the Trouble Free Pool (TFP) methodology.

It pulls data from three sources, feeds it to an AI that actually understands TFP chemistry, and delivers actionable advice like "dose 27 oz of liquid chlorine after sunset at 7:33 PM." Here's how it works and why I built it.

The Problem

Staying on top of pool chemistry requires juggling a lot of context: your latest test results, when you last added chemicals, what the weather's doing, how fast your chlorine is burning, and what your CYA level means for your FC target. I track everything in the Pool Math app, but checking the app, cross-referencing weather, and doing the mental math every day gets tedious — especially during DFW summers when you need to test and dose daily.

I wanted something that would do all of that thinking for me and just tell me what to do.

What I Get Every Morning

At 9am, a push notification hits my phone that looks like this:

🏊 Daily Pool Digest - Backyard
━━━━━━━━━━━━━━━━━━━━━━

🌡️ Temps
  Water: 66°F (24h avg: 66.3°F)
  Patio: 64.0°F

☀️ Weather
  Today: 62°/40° | clear | UV: 7 (high) | Wind: 17 mph | ☀7:41 AM-7:33 PM
  Tomorrow: 76°/47° | clear | UV: 7 (high) | Wind: 13 mph
  Day After: 78°/54° | partly cloudy | UV: 7 (high) | Wind: 24 mph

🧪 Chemistry
  FC: 4.5 ppm (tested 21h ago)
  pH: 7.5 (tested 21h ago)
  CC: 0.2 | TA: 70 | CH: 325 | CYA: 40

⚙️ Equipment
  Pump: On @ 50% | 268W
  Pool Heater: Off (pool 66°F → target 85°F)

✅ All readings in range. No action needed.

🤖 AI Recommendations (TFP)
Your FC was 4.5 ppm about 21 hours ago. Based on your recent
activity, your burn rate is 2.0-2.5 ppm/day and accelerating
as water temps climb...

Dose 27 oz of 10% liquid chlorine this evening after sunset.
This raises FC from ~3.5 to 6 ppm and builds a buffer for
tomorrow's UV 7 with 11 hours of sunshine...

Everything I need in one glance. If something is out of range, the notification comes through as high priority so it stands out.

The Data Pipeline

The system runs on N8N, a self-hosted workflow automation tool (think Zapier, but running on my own server). Here's what happens every morning:

1. Pool Math API — Pulls my latest test results and recent activity log from the Pool Math share API. This includes both test readings and chemical additions, which turns out to be critical for calculating FC burn rate. More on this below.

2. Home Assistant — My Hayward OmniLogic pool equipment is integrated with Home Assistant, so I pull real-time sensor data: water temperature, pump speed and power consumption, heater status, and ambient air temperature from a sensor on my patio.

3. Open-Meteo Weather API — A free weather API (no API key needed) that provides current conditions plus a 3-day forecast with UV index, wind speed, rain probability, sunshine hours, and sunrise/sunset times. The sunrise and sunset times are particularly useful — the AI uses them to recommend exactly when to dose chlorine.

4. Build the Digest — A code node assembles all the raw data into the formatted notification you see above, plus structured summaries that get sent to the AI.

5. AI Analysis — This is where it gets interesting.

Teaching AI the TFP Method

The biggest challenge with using AI for pool chemistry advice is that most AI models default to what I'd call "pool store advice" — recommending shock treatments, algaecide, phosphate removers, trichlor tabs, and other products that the TFP methodology explicitly advises against.

To solve this, I built a custom "skill" — essentially a reference document that gets loaded into the AI's context with every request. It contains:

• The complete FC/CYA table with minimum, target, and SLAM levels
• All TFP-approved chemicals (liquid chlorine, muriatic acid, baking soda, CYA, calcium chloride, borax) and explicit instructions to never recommend anything else
• Dosing calculations scaled to my specific pool volume (13,800 gallons)
• Weather impact factors specific to DFW — our intense summer UV, triple-digit heat, and storm patterns
• CSI calculation guidance for my plaster pool
• The full SLAM procedure
• My pool profile (plaster, manual liquid chlorine, TF-100 kit, no SWG)

With this context loaded, the AI doesn't guess. It knows that with CYA at 40, my minimum FC is 3 ppm and my target is 5 ppm. It knows to recommend liquid chlorine dosed after sunset. It calculates exact ounces for my pool. It recommends muriatic acid for pH, never dry acid on plaster. It will tell me to initiate SLAM if CC exceeds 0.5 ppm.

The AI I'm using is Claude (from Anthropic), running via their CLI tool called Claude Code in a Docker container on my server. The skill reference document gets appended to every request as system context, so Claude is always operating within TFP methodology.

The recentLogs Discovery

One thing I stumbled onto that I haven't seen documented anywhere: the Pool Math share API supports a recentLogs query parameter that returns your actual log entries — both test results and chemical additions.

By default, if you hit the share endpoint, the recentLogs array comes back empty:

GET https://api.poolmathapp.com/share/tfp-XXXXXX.json
→ recentLogs: []

But add ?recentLogs=10 and you get your last 10 entries:

GET https://api.poolmathapp.com/share/tfp-XXXXXX.json?recentLogs=10

This returns entries like:

Mar 11, 9:03 PM - Test: FC:4.5, pH:7.5, CC:0.2
Mar 10, 9:36 PM - Added 26.5 oz of 10% liquid chlorine
Mar 10, 9:35 PM - Test: FC:5.5, pH:7.4
Mar 9, 8:24 PM - Added 12.0 oz of 10% liquid chlorine
Mar 9, 8:19 PM - Test: FC:6.5, CC:0.5

This is a game-changer for the AI analysis because it can see the full picture — not just the latest snapshot, but the trend. It can calculate actual FC burn rate from real data and factor in chemical additions.

A few notes on the API: data is cached for 10 minutes, it's rate limited to about 1 request per minute, and sharing has to be enabled in the Pool Math app settings. No authentication token is needed.

FC Burn Rate and Prediction

With the recent logs data, the AI calculates the actual daily FC burn rate and predicts what FC will read at the next test. Here's a real example from my pool:

Observed FC burn rate:
Mar 10, 9:36 PM: FC had dropped to 5.5 ppm (1.7 ppm loss in 25 hours = 1.6 ppm/day)
Dosed 26.5 oz, FC jumped to ~7.0 ppm
Mar 11, 9:03 PM: FC dropped to 4.5 ppm (2.5 ppm loss in 23 hours = 2.6 ppm/day)

Burn rate is accelerating from 1.6 to 2.6 ppm/day as water temps climb.

Projection: If you dose 27 oz tonight to 6 ppm, expect 3.5-4.0 ppm by tomorrow evening.

This is the kind of analysis that would take me 10 minutes to do manually with a calculator and weather app. Now it just shows up on my phone while I'm drinking coffee.

I'm tracking prediction accuracy over time. Early results have been surprisingly close to actual readings.

The Weather Factor

One of the things that makes this system more useful than just checking Pool Math is the weather integration. The AI looks at the 3-day forecast and anticipates how it will affect chemistry:

• High UV with long sunshine hours — faster FC degradation, may need to dose above target to build a buffer
• Rain forecast — potential dilution, pH drop, organic contamination; sometimes better to hold off on dosing until after the storm
• Rising temperatures — accelerating chlorine consumption and increased algae risk
• High wind — debris and organics blown in, increased chlorine demand, pH drift from aeration
• Sunrise/sunset times — specific recommendations for when to test and when to dose

In DFW, where we can go from 65°F to 95°F in a couple weeks during spring, this forward-looking advice is really valuable. The AI catches things like "your burn rate will jump from 2 to 3+ ppm/day as temps climb to 78°F — dose heavier tonight to stay ahead."

What's Next

A few things I want to add:

• Prediction tracking — Log predicted vs actual FC at each test to measure and improve accuracy over time
• Feedback loop — Adjust recommendations based on whether I actually followed the previous day's advice
• Pump automation — Potentially integrate with my OmniLogic to adjust pump schedules based on weather forecasts

The Tech Stack

For anyone curious about the technical details:

• N8N — Self-hosted workflow automation, running in Docker
• Claude Code — Anthropic's CLI tool running in a Docker container with a custom TFP pool chemistry skill
• Pool Math API — The share endpoint with the recentLogs parameter
• Home Assistant — Hayward OmniLogic integration for real-time equipment and sensor data
• Open-Meteo — Free weather API with UV index, forecasts, and sunrise/sunset
• Ntfy — Self-hosted push notification server
• Cloudflare Tunnel — Secure external access to everything without opening ports

Everything runs on a single homelab server. The whole workflow executes in about 90 seconds — most of that is waiting for the AI to generate its analysis.

If you're into pool chemistry and homelabbing, this has been one of the most satisfying projects to build. Happy to answer questions about any piece of it.

This week it was my turn to make dinner for the family. and I wanted to do something a little bit better just for a random Monday night than a run of the mill meal. I wanted to have a full-on steak dinner with rosemary potatoes, Brussels sprouts, and a new salad that I'd never tried before. It was kind of ambitious. Two new recipes that I've never done before. Brussels sprouts I've tried before but did not do well at. But steaks, though, I was very confident in. I've grilled or sous-vided hundreds of steaks in my life. So that one I wasn't worried about. It was the rest of everything and coordinating it all and getting the schedule lined up perfectly so that we could eat at a decent time and also give my wife a break and not have to do anything.

So, I did a Google search and found some recipes for some side dishes and that's where I came up with the Brussels sprouts that I'm gonna try, the salad, and the rosemary potatoes. I knew that I wanted to try these recipes. And there was a lot of information to gather together, shopping list to build, and a schedule to put together for getting dinner ready. And on top of all this, I hadn't shopped for any of this yet. so I had to make a trip to the grocery store and get everything I needed and get started.

Now remember, this was a Monday and I didn't even start my trip to the grocery store until until four o'clock. And yet, I still had dinner on the table a little before 7pm and consensus was it was It was one of the best meals we'd had in a long time. And here's a secret. I used my AI assistant, Claude, to pull together the shopping list schedule and an overview of the recipe and guidelines into one simple to follow document that made pulling together the meal relatively easy for the scope of what I made. So, let's talk about it.

The Task I Actually Gave It

The primary AI tool that I use for almost everything is Clawed. it does a lot of things really well and since I I spend most of my professional life if dealing with coding and software development, it is the best tool for me to use. So I have a premium subscription to their model. Along with that, I've started using my own self-hosted note and to-do list app called Joplin, where all of my notes are synced together on any of my devices and it's on my HomeLab server, which we'll talk about another time in another post. It syncs it all together so that I can have notes on my MacBook, my Windows desktop, or my iPhone, phone and I can get to any of them at any of time.

Here's a screenshot of the actual prompt in the beginning of the chat session that I gave, just to give you an idea of the type of prompts that I use with my AI.

As you can see, I gave it quite a bit of information. I gave it the context of my family, the idea of what I wanted, and the three links that that I wanted to use and the expected output that I should get when they pulled it all together. I did make adjustments along the way. I subbed a few things out and had to change the plan. I originally wanted to cook everything on the grill, but because of a little bit of a time crunch when everything started taking a little bit longer to get going, specifically getting my grill up to temperature, I had to start cooking some of it inside in the oven and on the stove top. So I told AI about it and it adapted the schedule for me.

Normally if you're putting together a shopping list, a recipe and a schedule. It's kind of a tedious task. You gotta go to each page, skip through the ads, close the pop-ups, type out ingredients you want, figure out exactly where everything is. and it takes a lot of time and it's not fun for most of us. We don't really like pulling this all together. But what was great about using this particular prompt and this method and telling the AI exactly how it should give me the output I wanted, it was actually incredibly fast. and it had a shopping list and a plan for me that I was able to follow that helped me stay focused both when I went shopping and while I was prepping the meal.

Let me just point out also that it didn't really require this super powered AI engine to pull it all together in this deep magical way. What it actually required was attention to detail, organization and in a sense patience, which AI actually does a really good job at all of these. Take a look at the shopping list that it put together for me here in this screenshot.

If I'm honest, I would not have had my shopping list anywhere as close to as organized as the AI gave it to me. It put the items that I needed, not only in a checklist, but also in the section of the grocery store where I should expect to see them. So it made my shopping trip efficient and I didn't forget anything. I was able to check them off from my iPhone in my Joplin notebook. This was a really easy to use process.

What It Did Well

Was it perfect? No, but it was really, really good in most respects. Here's kind of a quick overview of some things that it did well:

• Fetched all three recipes without me having to open a single tab
• Scaled quantities for 5 people across three different recipe formats
• Noticed that the Brussels sprouts recipe was oven-only and added grill adaptation notes unprompted
• Built a cook timeline showing when to start each dish relative to serving time

What It Still Can't Do

One area where it didn't really give me the best result was the timeline for everything. just like life, things didn't quite go to plan. So the expected timeline of about an hour to an hour and a half of cooking and prepping It took almost two hours, which to be honest was not bad. We ate at a reasonable time, and it was definitely worth the wait. it really did a good job of pointing out also on the Brussels sprouts in particular, that I really shouldn't try to do those on the grill, that I should do those on the oven and it was worth it. The Brussels sprouts were really, really good, and I'm for sure going to do that recipe again. And surprisingly enough all three of my teenage kids, for the most part, liked the Brussels sprouts. So that's a big win.

Another thing I want to point out is that just like using AI for or building stuff in the software world, one of the key things that I learned is iteration is key when using AI. AI. Every now and then you're going to get a nearly perfect output or response the first first time you ask a question. But most of the time, you have to iterate three, four, or five times and get clarification, edit changes to make the output put exactly what you need or want. And with this dinner prep and shopping experience, I'll do a few things a little bit different. I'll probably tell the AI to buffer in at least an extra 15 to 20 20% extra time than it expects it to take because some things don't work. And honestly, I need to plan this out earlier than the afternoon of when I'm I'm gonna cook. The day before it would have been way smarter because there was things I could have done in the morning and started things a bit earlier and had dinner done an hour earlier than I did. But live and learn.

The biggest takeaway that I want to share about this is that AI doesn't have to be about coding or making images or funny short videos that you share on social media. using AI can give practical efficiency and assistance for routine daily things as simple as making a meal. I didn't ask it to teach me how to cook a perfect steak. I didn't ask it to tell me how to set the table or which exact spices to use to put on my steaks. These are things that I learned from experience in life. but putting together a list and having a document to help me stay focused and remember all the little things that are required to make a meal, it was incredibly helpful.

How to Try This Yourself

You don't need a special setup to replicated this yourself. Claude, ChatGPT, Gemini and similar tools can all do this kind of task. The key is being specific about your output format: "give me a markdown file," "scale to X servings," "I want a timeline." The more structure you ask for, the more useful the result. But for sure you want to paste URLs directly into the conversation rather than describing the recipes and, again, treat the output as a draft, not a final answer.

Finally, I just want to remind you that AI didn't make me a better cook. I still could have burnt the bacon, and I still could have burnt the Brussels sprouts. I could have over or undercooked the steaks. all those are things that I still had to do with intuition and life lessons. so AI is not going to replace everything in life. But AI did help me organize my prep and shopping so I could focus on just making a great meal for the family. and we all enjoyed it and a great evening and there weren't really any leftovers, which tells you how good the meal actually was.

There's a common saying ,"the numbers don't lie," but is that true? Well, kinda. If you're tracking stats consistently you've got an objective dataset that in an of itself doesn't lie. But what you do with that data or how you arrive at that data can be manipulated in a way to tell the narrative that you're wanting to tell. You can tell a story from true data that doesn't hold as true when you look at it in a broader context.

Let's look at an example with the current Covd-19 Pandemic. Here's a Tweet from Andy Slavitt on May 2:

A couple points first. I don't have the exact data sources that Mr. Slavitt was using for his numbers, so the exact daily count that I refer to in a moment won't match. My sources were Worldometer for USA and the New York State Department of Health Covid-19 Tracker for New York. With that out of the way, let's take a look at the Tweet and the data itself.

First, let's talk about the truth of the data. In a 7-day period the number of new cases in the USA outside of New York did increase, per his data source, by 17%. And at that growth rate we would see over 50,000 new cases by Memorial Day. The math adds up in and of itself. But why choose those two days in isolation? What happens if we look at other 7 day changes right around that time? Let's take a look at this table:

Again, my data source is probably a little different than Mr. Slavitt, so the exact numbers won't match. But let's jump forward a couple days to the 7-day period ending 5/3. If we use that date then the case numbers outside of New York actually grew by even more (19%) than the seven days he referenced. If we use that date and percentage growth than we'd have even more than 50,000 new cases by Memorial Day.

The problem with either of  these two dates is that it's not looking at those individual days in context. Using this same approach we could actually forecast the opposite effect. The seven days ending 4/30 saw a drop in cases outside NY by -6.2%. So using that number we'd project just over 20,000 new daily cases by Memorial Day. I used the exact same logic and a date only two day prior to produce a result that was 60% lower. That's the danger of using such a small data point out of context to make proclamations. You really need to look at the context and a bigger range of days to get a better feel for how things are going.

I think a more useful thing to do is compare the seven day average between two days. That will do a better job normalizing the numbers. On 4/29 the average was 24,128 outside NY and it was 25,463 on 5/5. That's a 5.5% growth in a week. So, again, using just that logic and a small dataset that projects new daily cases around 30,000 at Memorial Day.

But what does the trend actually look like? Here's a chart to help show the numbers:

For the last three days the USA has seen daily cases outside NY at or below the seven day average. Part of the reason that the seven day trend is going up is the spike in cases on 5/1. That number will make the average grow, for sure. But a very small downward trend might be starting now. We will see soon.

So, what's the takeaway from all this? Simply this. It's nearly impossible to know what the future will bring with this virus. If you want to paint a rosy picture you can find the data to do that. If you can only see doom and gloom you can find the data to support that opinion.

My personal read on the data in a broader sense is that things will start to greatly improve between now and Memorial Day. I am generally an optimist, so this is my default position. But I also think the data is starting to point that direction. Most states are seeing an overall downward trend in most every key metric. Time will tell.

Stay smart and stay safe everyone. This too shall pass. Hopefully soon.

I'm a geek and a sports nut. So, that makes baseball one of my favorite things. Ever since I was in Pony League I was tracking stats. I loved knowing what my batting average and slugging percentage were. This was before advanced analytics, Moneyball, and sabermetrics were commonplace. I would also spend hours every week reading the MLB player stats on the back of their Topps Baseball Cards and in the box scores in the Sports section of the newspaper.

Nowadays it's much easier to track player stats. Thanks to the free website from Baseball-Reference.com anyone can find just about any stat that they can dream of. Many of the stats provide useful insights while others are just fun to dig into. On top of that, visualization tools like Power BI allows anyone to more easily see data in ways that the numbers alone can't really do justice to. Let's take a look at one example: Wins Against Replacement (WAR).

WAR is a relatively new stat that tries to give more insight into how valuable a player truly is to his team. It looks past batting average and tries to determine the full impact a player has had in his team's wins. The Texas Rangers just played their 81st game of the 2019 season and they have far exceeded my and just about everyone else's expectations. One player in particular has had a breakthrough season: Joey Gallo.

Gallo hit 40+ HRs in each of his two previous seasons, so you might wonder why this season could possibly be considered a breakthrough season. Well, his highest batting average in any previous season was .209 and he struck out nearly 200 times both of full seasons in 2017 and 2018. This year, however, he's hitting .279 and has only struck out 80 times, although he has missed about 25 games due to injury.

Gallo's injuries are quite interesting when you look at the league leaders in WAR. At the midway point only 16 players have a WAR of at least 3.0, and Gallo is one of them. Take a look at the following chart:

Gallo is one of the outliers. He's WAAAAYYYYY down at the bottom left. There are no other players in the league that have a WAR of 3.0+ who have played in less than 72 games, but Gallo has accomplished this in only 52 games with a 3.1 WAR. The average number of games played for the players with 3.0+ WAR is 76. If Gallo had played that many games in 2019 his WAR would likely be around 4.5, which would put him behind only Cody Bellinger (6.6) and Mike Trout (5.2).  (Note: this article was written before Baseball-Reference had updated their data to include June 27th's games in which he hit two home runs, so his WAR will likely be higher when the data is updated).

Taking a look at the leaders' current WAR Per Game and extrapolating that over a full 162 games would make the leaders look like this:

So, where does all this data lead us? First, it's a little disappointing that Gallo was not voted as a starter in the 2019 All-Star Game. Understandable for sure, as there are still a few Rangers fans who dislike Gallo because he doesn't bunt to get on base against the shift. But most importantly this data shows that Gallo is a rising star. Looking at his impact on the Rangers season thus far it's undeniable that he's the most impactful player on the team's roster. This doesn't guarantee that that second half of the season will be as good as the first half. But it sure has been a fun ride so far.

If you'd like to take a look at the interactive report I made in Power BI you can check out it using this link.