That is why every entry keeps the same three records: technical question, working answer, and honest limit.
OPEN LIMIT — STOPS AT 72%
EXPERIMENT SURFACE04 / 06
04 / Frontend systems experiment
Calm the weather data.
DATA FIELD81 / 81
SCAN · OPENWEATHER TR / EN · THEME STATE
81PROVINCES · ONE SURFACE REACT 19 · LEAFLET
05 / Interactive learning tool
Trace failure layer by layer.
SIGNAL PATH07 LAYERS
07APPLICATION
06PRESENTATION
05SESSION
04TRANSPORT
03NETWORK
02DATA LINK
01PHYSICAL
TRACE: —
10FAILURE SCENARIOS OSI · 7 LAYERS
06 / Model-evaluation retrospective
Question the test design, not the score.
VALIDATION AUDITR² ≈ .99 / REPORTED
R²≈.99NOT PROOFRANDOM 80/20 SPLIT · REPORTED
LEAK / 01
camera_price_ratiocamera_mp ÷ price
TARGET → INPUT
MEMORY / 02
model_encodedknown identity repeats across the split
IDENTITY MEMORY
UNSEEN-MODEL GROUP HOLDOUTNOT RUN
NEXT TESTRemove the target-derived feature · split by model group
0404
Frontend systems experiment
Türkiye Weather Dashboard
DATA FIELD81 / 81
SCAN · OPENWEATHER TR / EN · THEME STATE
81PROVINCES · ONE SURFACE REACT 19 · LEAFLET
QUESTION
How can live data for 81 provinces, a map, and user preferences remain calm on one responsive surface?
WORKING ANSWER
React 19, TypeScript, Leaflet, OpenWeather, two languages, and a weather-aware theme work together.
OPEN LIMIT
API keys and provider limits keep the live-data layer within client-demo boundaries.
81 provinces
React 19
Leaflet
TR / EN
0505
Interactive learning tool
Network Troubleshooting Guide
SIGNAL PATH07 LAYERS
07APPLICATION
06PRESENTATION
05SESSION
04TRANSPORT
03NETWORK
02DATA LINK
01PHYSICAL
TRACE: —
10FAILURE SCENARIOS OSI · 7 LAYERS
QUESTION
Can the OSI model and ten common failures become inspectable scenarios instead of a memorized list?
WORKING ANSWER
A working learning surface built with vanilla web technologies, GSAP, bilingual content, and selected network simulations.
OPEN LIMIT
The content is strong; its loading rhythm and visual system need revision before this becomes a main case study.
7 layers
10 scenarios
GSAP
TR / EN
0606
Model-evaluation retrospective
Phone Price Predictor
VALIDATION AUDITR² ≈ .99 / REPORTED
R²≈.99NOT PROOFRANDOM 80/20 SPLIT · REPORTED
LEAK / 01
camera_price_ratiocamera_mp ÷ price
TARGET → INPUT
MEMORY / 02
model_encodedknown identity repeats across the split
IDENTITY MEMORY
UNSEEN-MODEL GROUP HOLDOUTNOT RUN
NEXT TESTRemove the target-derived feature · split by model group
QUESTION
What does an R² near .99 mean when a target-derived price feature enters the model?
WORKING ANSWER
A working prototype built from roughly 290 Trendyol records, a random 80/20 split, Gradient Boosting, and a Streamlit interface.
OPEN LIMIT
camera_price_ratio leaks the target price into the inputs, while model identity repeats across the random split. The reported score is not evidence of generalization.
≈290 records
R² ≈ .99
Target leakage
Group split needed
LAB / NEXT
Products live in selected work. Open questions live here.