Full-stack analytics leader turned entrepreneur Co-Founder & CTO @policyquail - compare auto policies Co-Founder of Falcon Media - digital marketing agency
NYC
- Tweets 99
- Following 188
- Followers 32
- Likes 98
Have been testing Grok 4.5 recently and it's quickly becoming my daily driver for code implementation.
Love the balance of speed and intelligence at the low cost. Have been using GPT 4.6 Sol xhigh for building the plan/orchestration, then handing off to Grok 4.5 for execution.
Once execution is completed, pass back to GPT 4.6 for a pass through to find any potential bugs and comment on the PR. Sometimes have to go through a few cycles of this, but it works really well all together.
Still playing around with the best workflow, but really enjoying this process for quickly implementing features.
2
8
55
What all these new releases of AI models have reinforced in me is that it's not the tools that you use that matter. It's making sure you're getting done what needs to get done for your company to be successful.
Doesn't matter if you're not using the latest model or the most efficient workflow. You just need to be making progress on your goals.
Use tools that you enjoy and that actually make you productive. Don't need to be constantly chasing the next big thing.
1
34
Nothing beats a good dashboard. Just give me the metrics I need. Don't need a chat interface to tell me what I'm suppose to be looking at
1
48
A while back I noticed I was actually enjoying building again. The shift: tracking what made productive days feel different from wasted ones.
Then isolating the patterns. Then repeating them.
Here is what stuck:
1) I work in focused blocks on one project at a time. Scattered days killed momentum.
Single focus kept things shipping.
2) I automate the repetitive stuff first. When I cleared the noise early, everything downstream moved faster.
3) I build tools that solve real problems (QuoteFii started as "why is comparing insurance so painful"). Visible user impact made the work feel worth it.
4) I review data before making any decision. Gut instinct plus a dashboard beats gut instinct alone every time.
The pattern was simple: notice what made a good day, remove what made a bad one, repeat until it felt like a system.
TLDR: Ship one thing. Automate the friction.
Use data. Ignore the rest.
1
38
Recently passed 3k contributions for the year and we're not even halfway through 2026.
My previous record was ~700 contributions for a given year. Claude code and Codex have made it way easier and more fun to ship 🚤
1
61
I built a pipeline that turns public insurance regulator filings into clean rate datasets.
The work starts with company submitted prices that regulators publish. They are closer to the source than aggregator averages, but they still arrive in scattered public formats. Before I use them, the pipeline normalizes the filings into one schema.
Then the pipeline isolates one variable at a time. If I want to understand one factor, every other attribute has to stay fixed. Otherwise the number is just a blend of several changes.
After that, the system compares the pattern across independent jurisdictions and grades the evidence. If the sources agree cleanly, the claim can move forward. If the data is muddy, the article has to leave that claim out.
The goal is fewer, better numbers, backed by a dataset where a rate claim only ships when the evidence is good enough to survive being checked.
1
3
42
The evidence gate is where most of the judgment lives.
A row can be real and still be unusable. If the variable isolation is weak or the source set is too thin, the system should hold the claim back instead of laundering it into a confident sentence.
1
24
I built statistical modeling into my Google Ads reporting because the normal dashboard was too good at sounding certain.
A daily row has all the official looking parts. Clicks, conversions, cost, value, conversion rate. If any of them move, the dashboard presents the movement like a fact. Recent data may still be settling, weekend samples can be thin, and the auction underneath the account can change while I am staring at yesterday's number.
The math was the easy part. The annoying part was making the report admit when the evidence was weak. CPC, conversion rate, and average conversion value each get compared against a baseline as distributions. The report can call a driver STABLE, ELEVATED, DEPRESSED, or UNCERTAIN, and UNCERTAIN is a real output rather than a blank spot.
That changed my morning read in a useful way. I still see movement, but the report makes me ask whether the movement earned a reaction. Sometimes a driver moves far enough, with enough evidence, that it deserves attention. More often, the current sample is thin, attribution is still settling, or the driver moved inside a wide credible interval.
That answer is quieter than a red dashboard cell. It is also usually the answer I need. The tool has kept me from turning ordinary noise into account work more times than I can count.
1
42
I built my own internal marketing dashboard after getting tired of stitching together BI tools, spreadsheets, and reports that never quite answered the question I had that morning.
The first win was obvious: I avoided at least $300/month in BI software before counting setup time, dashboard maintenance, or the usual report-disagreement tax.
The bigger win was control.
I can decide how fresh each dataset needs to be, which metrics deserve confidence flags, what gets cached, what gets ignored, and how paid data connects back to actual business outcomes.
A generic dashboard can show me yesterday’s numbers.
My dashboard can tell me whether the data is settled enough to trust, whether a movement looks like signal or noise, and which part of the system needs attention.
I used to think the goal was a better dashboard.
The more I build, the more I think the real goal is a marketing operating system that understands the business better than rented software ever could.
1
14
Dashboards make small samples look confident. In Google Search ads, 24% CR (6/25 clicks) sits there wearing the same font as 24% CR (60/250 clicks). That is how noise gets promoted. I want the denominator next to the rate before the rate gets a vote
2
14
A second tool keeps me honest when something does move. A change history auditor tells me whether I changed something or the algorithm did. Most panic comes from blaming a shift that nobody actually caused.
1
12