Kacper Michalik

Software Engineer @ Screen Studio (#1 PH 2025) · AI/ML · Python · TypeScript · Product Engineer

[Adam Pietrasiak @pie6k]

Adding macOS 27 inspired sidebar glow to https://t.co/0AoQbS3QmQ

It is watching the canvas and generating glow in realtime using typegpu.

Had a lot of fun cooking it.

[Kacper Michalik @casp3ro_]

How do you hire really matching engineers at a startup without an HR team? No recruiters, no process.
Genuinely curious what works for you.

[Kacper Michalik @casp3ro_]

Retention query from scratch in DuckDB, pure SQL.

[Kacper Michalik @casp3ro_]

Back to roots today, instead of nice Prisma queries, going raw in DuckDB.

[Kacper Michalik @casp3ro_]

Great day at @appjsconf - really enjoyed the TypeGPU talk.
See you tomorrow at @screenstudio booth.

[Kacper Michalik @casp3ro_]

Redesigned view of created trips in TravPlanner

[Kacper Michalik @casp3ro_]

Polishing animations with Framer Motion.
Feels more smooth and fluid.

[Kacper Michalik @casp3ro_]

Refactored single analysis flow to multi-agent approach - orchestrate in a 3-step pipeline: understanding data → creating charts → generating insights.

[Kacper Michalik @casp3ro_]

Enhancing RAG chat-based experience

[Andrej Karpathy @karpathy]

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[Kacper Michalik @casp3ro_]

Back to (very)old days as a designer.
Except now I design through code.

[Kacper Michalik @casp3ro_]

used my own tool to analyze the dataset I’m training a model on.

Sierra Graph surfaced what matters:
→ OverallQual is the strongest pricing driver
→ neighborhood premium can triple the base price
→ 3-car garage adds a significant jump vs 2-car

#MachineLearning #DataVisualization

[Kacper Michalik @casp3ro_]

Second Kaggle submission - House Prices regression.

key factors:
→ train on log(SalePrice), not raw price
→ engineer HouseAge, TotalSF, TotalBath features
→ fastai tabular_learner with embeddings for categoricals
→ layers=[1000, 500, 250]

current score: 15,161 RMSE

more optimization to come.

https://t.co/SaGXx6mFmc

[Adam Pietrasiak @pie6k]

25,000 people used @screenstudio last month.

We’re still a team of 5. No investors. Every hire came directly from our revenue.

There’s something humbling about that. But also something that’s been holding us back.

It’s time to think bigger.

[Kacper Michalik @casp3ro_]

Delta Lake stores data as Parquet files.
The _delta_log directory is what makes it a lakehouse.

each JSON file = one atomic commit:
→ version 0: initial write
→ version 1–6: subsequent pipeline runs

this log gives you:
- ACID transactions
- time travel (read any version)
- schema evolution
- upserts via MERGE

same Parquet files. completely different guarantees.

#DataEngineering #DeltaLake

[Kacper Michalik @casp3ro_]

304k car listings. 41 makes. 50 states. $14,995 median.

all from a pipeline I built in a few days: Spark → Delta Lake → dbt → DuckDB → Streamlit.

#DataEngineering

[Kacper Michalik @casp3ro_]

Medallion architecture explained with real data:

🟤 Bronze - raw Delta Lake on MinIO. immutable. never touched. 49 parquet files, 1.5GB, _delta_log for ACID transactions.

⚪ Silver - Spark cleans it. removes duplicates, filters bad prices ($500–$200k), normalizes strings.

🟡 Gold - dbt aggregates it. median price by make, depreciation curves, state heatmap.

#DataEngineering #DeltaLake #dbt #Medallion

[Dawid Wilewski @AL8ATROS]

One does not simply walk into production.

Testing new @screenstudio Teams with @casp3ro_