JimmyDans🎯

And yes, I asked Claude to create a posts folder locally, search through all of our previous chats, highlight the stuff we've built, and create daily journal md files of what seemed important. All I am is an AI manager now.

I had a meeting on Friday, bc I had delivered some pretty amazing stuff over the week, a #streamlit dashboard, 3 actually, some incredible analysis over the cost of a 1% increase in queue decision rates, had #claude build a nice PowerPoint out of it too, and some other stuff. The folks in the meeting said I must be a spokesperson for #anthropic the way I talk about #cursor and #claude. I assure you I am not. Wish I was honestly

I mostly keep coming back to the idea that there's no way to predict what a data job, no matter how senior I am, will be like in a year. Maybe less. All I know to do is get as knowledgeable as one can be on how to manage #claude, sell my house, maybe rent for a while, and buy low

Second week So I have to build a new very large awful doc for a crypto model from a vendor for a team called model risk management. I know nothing about this crypto model, mrm, governance, any of that. But it's on me to build the documentation report. The api info is all behind a login for the crypto vendor and a diff login for this risk model they have. So I have to go scrape it all manually. So i do, tossed it all in large text files, 7k lines of garbage, included the model template I'm suppose to use, some previous model docs that got approved, put it all in a folder and told Claude go build me this doc. I even connected the slack and Google drive mcp servers and said go read through these channels and look for any Google docs around this topic. In about 5 minutes it built a 26 page model documentation for this risk model. I asked it to call out things in red that it didn't know, whenever it makes changes from new prompts, mark in blue. There's even a section that requires some actual crypto metrics, I told Claude you know where our #crypto data is, go get those metrics bro. It did. I've done a few of these mrm docs, they take months, 3-6 easy. I still need to validate, but that's all I am these days. #ai #claude #fraud #DataScientist #cursor

Day 7 continued Real scenario from today: My boss has lots of random queries she puts in Google Docs, snowflake worksheets, Google Sheets - everywhere, and says "I need you to go through these and create xyzzy". It takes weeks. Part of that is getting it deployed, other part is making her code actually work. She had a login model with thousands of lines of code, Claude reconfigured the entire thing, bunch of cells, and had it up and running within an hour. Her code prior took half a day just to create all the tables used in the final. I told her, look I don't need a meeting, I just need you to send me bullet points and I'll take it from there. Quite a different conversation than a month ago. #DataQuality #Analytics #EnterpriseData #SQL #AILimitations

Day 6 continued After the AI built 130 features, I said: "double it." It added: - High-dollar pattern detection - Cross-channel velocity features - Historical behavioral baselines - Risk score interactions Then it found a column that doesn't exist in the table it was querying. Fixed it. Found another table in a different schema that had the right data. Used that instead. The AI self-corrected mid-build. That's the part nobody talks about. I often say things like "add anything else you see fit", "double check your work" etc. #AIAgent #SelfCorrection #DataEngineering #ML #BuildInPublic

Day 7. Asked the AI to build a fraud detection feature sweep. It created a notebook with 130+ features across: - Device fingerprinting - Geolocation & IP behavior - Transaction patterns (P2P, wires, cards) - Profile change signals - Third-party model scores - Behavioral anomalies 42 cells. 19 sections. 10-minute query guard so nothing runs forever. It's actually incredible. Kind of took my breath away. Haven't really gotten it back. This would have been a full sprint for a data scientist. We did it in an afternoon. #MachineLearning #FeatureEngineering #FraudDetection #AI #DataScience

Day 6 continued PSA for anyone building Streamlit apps with AI: NEVER put `st.progress()`, `st.spinner()`, or `st.write()` inside a `@st.cache_data` function. Streamlit replays cached return values but NOT UI calls. You'll get ghost text, broken layouts, and text bleeding into the status bar. Cache the data. Orchestrate the UI outside. Learned this the hard way. Made it a permanent rule. #Streamlit #Python #BugFix #WebDev #DataApps

Day 6 continued Built a complete post-deployment ATO model monitoring system: - Comparative benchmarks vs pre-launch - Verification volume tracking - Fraud trend breakdowns - Operational complaint analysis - Executive chart section Notebook → Streamlit app → Download buttons (HTML, PNG, CSV, Excel) All in one session. With an AI. Only 1400 lines of code this time. I keep waiting for the catch. Haven't found it yet. #Analytics #ModelMonitoring #Streamlit #Python #DataScience

Day 5 continued It dawned on me that my snowflake clone never has to actually delete... regardless of the 30 day drop policy my company set. I had Claude build a clone recreation/rotation script which creates a new clone, saves the custom tables in its stage, lets the old clone get dropped ofter day 30, recreates the same name clone 30 days later with the same custom tables etc etc. little confusing, but it will look like the clone never deletes. #Snowflake tasks - had no idea. #DataGovernance #BestPractices #SafetyFirst #SQL #EnterpriseAI

Day 5 continued Set up SQLFluff in cursor with a keyboard shortcut today. AI built: - A utility notebook that auto-detects SQL vs Python vs JSON - VS Code tasks for format-on-save - Keybindings for one-touch formatting I use to very much dislike sqlfluff, but that was b/c it was forced on me to pass go for dbt models. then I found the dbt ignore file ;) #SQLFluff #CodeFormatting #DeveloperTools #SQL #Productivity

Day 5. Started with "add HTML export to the dashboard." Ended with: - Run history saved to the database - Shareable analysis sessions - Dispute data integration - Heatmaps, funnels, and burst charts - Full self-contained HTML reports One feature request turned into six. The AI just... kept going. And everything worked. I presented this streamline dashboard this week - seems like it was a hit. I think it's pretty cool at least lol. I've never shipped this fast in my career. #Streamlit #DataApps #BuildInPublic #Productivity #AIAgent

Day 4 continued Discovered "rules" — persistent instructions the AI follows every conversation. Made ones for: - Always use pre-aggregated tables when available - Follow specific app patterns for Streamlit - Add confidence thresholds after every query produced -Add context and tldr descriptions -Keep data frames small, it's all in-memory Now every AI session starts with my team's best practices baked in. No drift. #DeveloperProductivity #AIGuidance #BestPractices #CodeStandards #CursorAI

Day 4 continued Asked Claude to investigate why a fraud model's metrics looked off (low AUC, poor recall). It diagnosed: deduplication issues, misaligned time windows, and inconsistent fraud labels across data sources. Then it built a standalone metrics module with multi-threshold AUC curves, lift charts, and proper cohort handling. MRM might approve it now. We'll see Sometimes the AI's best work isn't writing code. It's finding why your existing code is lying to you. #MachineLearning #FraudDetection #ModelValidation #DataScience #AIAssistant

Day 4 continued Had a login-event processing pipeline notebook with 6 slow queries. The AI identified: - 4 redundant full-table scans - 1 hardcoded IP list that could be a lookup table - 1 join causing a 10x row fan-out It fixed all four I approved in one pass. Like having a senior data engineer review your work at 2 PM on a Saturday. #CodeReview #DataPipelines #SQLOptimization #AIAgent #FinTech

Day 4. My Streamlit fraud investigation app was taking 90 minutes to run, or timing out. Asked the AI to optimize it. It replaced raw table joins with pre-aggregated summaries, added date-range filters, and eliminated redundant scans. and 3400 lines of code later. Result: 20 minutes. Same output. 4.5x faster. The AI didn't guess. It analyzed the query patterns and found the bottlenecks. #Performance #SQL #DataEngineering #Optimization #Streamlit

Back posting the last few weeks of using Cursor and Claude. Enjoy

Day 3 continued Real talk: the hardest part of connecting AI to cloud services isn't the AI. It's credentials. SSO tokens, refresh scripts, expired sessions, enterprise auth flows — spent more time on authentication than on any actual analysis today. Enterprise AI adoption is 80% infra, 20% intelligence. #EnterpriseAI #DevOps #CloudSecurity #RealTalk #TechDebt

Day 3 continued Here's a concept that changed how I work with AI: "Skills" — markdown files that teach the AI agent how to do specific tasks in YOUR environment. I created one for cloud storage operations. Now every future conversation starts with that context. No re-explaining. It's like onboarding a new team member once and having them remember forever. #AIProductivity #DeveloperExperience #KnowledgeManagement #CursorAI #TechTips

Day 3. I asked the AI to set up a connection to cloud storage (S3). It didn't just configure the connection. It: - Created a virtual environment - Wrote a credential refresh helper script - Built a reusable "skill" doc so it remembers how to do this next time - Set default bucket paths An AI that builds its own institutional memory. That's the part that surprised me. #AWS #S3 #CloudComputing #AI #DevOps #BuildInPublic