ozan

not looking for a reason

TOPSET:

[ozan @ozanc404]

Software is evolving with AI, and I believe people who can build systems end to end will become more valuable.

Instead of staying in one field, I’m setting out to see the whole picture. I’ll share my journey and everything I learn here.

#minusonetozero

[ozan @ozanc404]

@JotaTerrasa Thank you for your suggestions! I’ll take a closer look and get back to you.

[ozan @ozanc404]

Another topic I’m focusing on so I can build bigger projects later: embeddings. Here’s the flow of the document search system I built, step by step. #RAG #embeddings

[ozan @ozanc404]

@JotaTerrasa Semantic splitting retrieved about 53% fewer tokens on average in the top 5 results, but it was less successful at finding the correct answers.

Now, the question I'm really interested in is: Can I design a hybrid approach that combines the strengths of both methods?

[ozan @ozanc404]

@JotaTerrasa The results were:

Semantic splitting (90th percentile): Successfully retrieved the correct information for 63.3% of the questions.

Fixed-size splitting (500 tokens): Successfully retrieved the correct information for 96.7% of the questions.

[ozan @ozanc404]

Next up: saving the vectors to a vector database and running searches on it. I’ll write about that part next. #RAG #VectorDB

[ozan @ozanc404]

I run the same model twice. First to decide where to split. Second to embed the final chunks so they can be searched. Those vectors are stored in Milvus together with the chunk text. #VectorDB #Milvus

[ozan @ozanc404]

Also in 2.6.0: 10 languages, plate and dumbbell drawings under the weight, strength charts with an estimated 1RM trend per exercise, set notes and new widgets. The run hub on Home suggests your next run.

[ozan @ozanc404]

Topset 2.6.0: GPS run tracking, in the same app as your lifting log.

Live map, pace, distance and time. It auto-pauses when you stop and resumes when you move. Free, no ads, no sign-up.

https://t.co/v1SmrBgLNX

[ozan @ozanc404]

I derive metrics from raw location samples via pure functions. Live screen, summary, records and tests share one engine.

Auto-pause uses hysteresis, so it doesn't flicker at a traffic light. The active run is saved to disk and comes back paused if the app dies.

#SwiftUI

[ozan @ozanc404]

Also in 2.6.0: 10 languages, plate and dumbbell drawings under the weight, strength charts with an estimated 1RM trend per exercise, set notes and new widgets. The run hub on Home suggests your next run.

[ozan @ozanc404]

I derive metrics from raw location samples via pure functions. Live screen, summary, records and tests share one engine.

Auto-pause uses hysteresis, so it doesn't flicker at a traffic light. The active run is saved to disk and comes back paused if the app dies.

#SwiftUI

[ozan @ozanc404]

3 object types, all sqlite-backed: room (one per game), join guard (counts wrong code tries per device and ip), moderation (keeps reports). 0 runtime dependencies.

echo rtt from türkiye to the live server: p50 68 ms, p95 83 ms.

#buildinpublic #spacegoat

[ozan @ozanc404]

spacegoat's app is flutter, picked only for cross-platform. the server speaks plain websocket, so the app needs no sdk, just dart's built-in WebSocket.

the cost: no ready login or admin panel. i wrote room logic, rejoining and device token checks myself.

#flutter #spacegoat

[ozan @ozanc404]

each spacegoat room is a cloudflare durable object running at $0. part of why:

sockets stay open while the object sleeps (websocket hibernation), and sleep isn't billed. setWebSocketAutoResponse answers the phones' 10s "ping" without waking the room.

#cloudflare #spacegoat

[ozan @ozanc404]

🏃 Running & walking tracking is coming to Topset!

Track your route, distance, pace, calories, and more.

P.S. I had to walk for testing purposes, so please don’t judge my stats. 😂

https://t.co/yZZoyoThXX

#buildinpublic #indiedev