I wanted to get better at Dota 2. Not by watching guides — by actually playing. I tried Keenplay but it felt too rigid: no overlay, no model choice, no way to customize it to my playstyle. I'm a developer. So I just built my own.
What it does
It reads your live game state via Dota 2's GSI (Game State Integration) — the game literally sends JSON with your hero, items, and map state to a local server every few seconds. My app catches that, runs it through a RAG pipeline, and shows tactical advice in an overlay while you play.
You can also type questions mid-game: press F10, ask "should I fight now?", get an answer without alt-tabbing.
The part that actually took time
The hardest thing wasn't the overlay or the GSI integration. It was prompts for generating knowledge base chunks. Getting the LLM to produce actually useful, structured Dota knowledge — not generic garbage — took way more iteration than I expected.
Also: DeepSeek wins on price/quality for this kind of task. Claude and Gemini are great but expensive. Most cheaper models just output nonsense. DeepSeek hits the sweet spot.
Does it actually work?
Yeah. That's the part I'm most proud of — it gives real, relevant advice. Not hallucinated builds. Not generic tips. Actual context-aware suggestions based on what's happening in your game right now.
Full RAG pipeline, BERT embeddings, vector search — all in Go, all open source.
Top comments (0)