How I work, and what it cost to learn Nasıl çalışıyorum, ve bunu öğrenmenin bedeli
Four rules, each one learned by getting a number wrong first, and the table of every measurement that changed a decision. The per-project engineering detail lives on each project's own page. Dört kural — her biri önce bir sayıyı yanlış ölçerek öğrenildi — ve bir kararı değiştiren her ölçümün tablosu. Projeye özel mühendislik detayı her projenin kendi sayfasında.
Four rulesDört kural
EACH ONE LEARNED THE HARD WAYMeasure before you buildKurmadan önce ölç
The plan assumed ±35° of plate angle. I labelled 272 real plates, measured about 7°, and changed the generator instead of the measurement. Plan ±35° plaka açısı varsayıyordu. 272 gerçek plaka etiketleyip yaklaşık 7° ölçtüm ve ölçümü değil üreteci değiştirdim.
Check the number that flatters youHoşuna giden sayıyı kontrol et
I concluded sharpness does not predict night readability. It does, from 1% to 62%. My first sample only covered the flat end of the range. Gecede keskinliğin okunabilirliği öngörmediği sonucuna varmıştım. Öngörüyor, %1'den %62'ye. İlk örneklemim aralığın sadece düz ucunu kapsıyordu.
Publish what fails tooBozulanı da yayımla
Every repo has a section on where it breaks. The trafikisaret demo keeps a night frame with one sign found, because that is how the model performs there. Her reponun nerede kırıldığını anlatan bölümü var. trafikisaret demosunda tek işaretin bulunduğu gece karesi duruyor, çünkü model orada böyle çalışıyor.
Write down what you could not doYapamadığını da yaz
The plate plan needed a Jetson board I do not have, so the edge-latency phase is marked out of scope rather than estimated. Plaka planı bende olmayan bir Jetson kartı gerektiriyordu; uç cihaz gecikme fazı tahmin edilmek yerine kapsam dışı işaretlendi.
Findings that changed a decision Bir kararı değiştiren bulgular
| Project | Assumed | Measured | What changed |
|---|---|---|---|
| trafikisaret | Bigger inference resolution is better | recall 0.379 @640 → 0.724 @960 → 0.448 @1280 | Inference runs at 960 px. There is an optimum, and 1280 measuring worse is what proves it. |
| trafikisaret | Leakage inflates a score | mAP50 0.694 dirty → 0.726 clean | Leakage made the number meaningless, not higher. Locked behind a test; the dirty run was kept, not deleted. |
| plakatanima | ±35° horizontal yaw | ~7° over 272 labelled plates | The generator was recalibrated against the labels instead of the plan. |
| plakatanima | Sharpness does not predict night readability | 1% → 62% across the full range | The first sample only covered the flat end of a sorted queue. The conclusion was withdrawn. |
| smart-city | Busiest zone is the critical one | JFK removal: 1 → 16 components | Volume and structural role are different questions. Betweenness catches what PageRank ranks 17th. |
| payflow | The chained-split example proves per-step rounding | 30% then 18%: 5.73 either way → 3% then 8%: 8.91 vs 8.92 | The example agreed with the rule without being able to distinguish it from the alternative. Replaced with percentages where the two rules disagree. |
| cse331 | $signed(a) >>> shamt is an arithmetic shift | sra 0x80000001 by 4: 0x08000000 , expected 0xF8000000 | Signedness in Verilog belongs to the whole expression: the unsigned arms of the conditional coerced it back to a logical shift. Moved into its own assignment. |
| floorplan-parser | Four radii per gap are enough for the door-arc template | one door: 0.62 @91 px → 1.00 @88 px | The radius step was about 7 px, more than twice the 3 px tolerance. Finer radii for door-sized arcs found 32 more doors on the development plans. |
| ada-llm-gnn | Message passing along door edges can tell whether a room is reachable | crossable small room, wrong: 28 / 40 → 4 / 40 | Whether a room too small to turn in can be crossed depends on the walls of two of its doors at once, which one door edge never sees. The route agent moved to a transit graph whose edge crosses a room and carries both walls. |
The full record Tam kayıt
Education, work experience, project experience and skills on one page. Eğitim, iş deneyimi, proje deneyimi ve yetkinlikler tek sayfada.