- Private
- Client
- In progress
- 2026
LLM coach for a competitive gaming platform
read-only · the LLM never calculates a number
Source 1
6 tools
6 read-only tools; the LLM never calculates a number
- How it was measuredBy design: every figure is computed in TypeScript tools against the ranking ledger; the model only chooses which tool to call
- Keep in mindProject in progress; tool choice is benchmarked across LLMs and answer wording is still generated
The problem
LLM coach for a competitive gaming platform · before
Points without reasons
A player sees "1,340 points, Gold" and doesn't know how they got there or what to do next.
One wrong number
On a platform whose value is trusted rankings, one invented number does more harm than no assistant.
The answer exists
But every point is an itemised ledger record, so "why?" has a real answer to retrieve.
The goal
LLM coach for a competitive gaming platform · after
Explains rankings
An assistant that explains a player's ranking in French, Arabic or English.
Numbers from tools
Every number comes from a read-only tool call into the real ranking services.
Never calculates
The model chooses the tool; it never calculates.
How it works
LLM coach for a competitive gaming platform
The player's question goes to a single tool-loop agent with a step limit. The agent picks from six typed, read-only tools that call the platform's existing ranking services; the player's identity is attached by the server, never by the model; the answer streams back in the player's language.
- Player · FR / AR / EN
- Rate limit + budget
- Tool-loop agent
- 6 read-only tools
- Session identity
- Ranking services · ledger
- Step 1 of 4 A player asks; the server applies a rate limit and a token budget before anything reaches the model.
- Step 2 of 4 One agent with a step limit: almost every question is a single lookup.
- Step 3 of 4 Six typed tools wrap the existing ranking services; any arithmetic runs in TypeScript, never in the model.
- Step 4 of 4 The player's identity is added by the server, never by the model, and no tool can write.
Next project
62 % Generative Engine Optimization ← Previous: CBAM compliance agentSources
6 tools
6 read-only tools; the LLM never calculates a number
- How it was measuredBy design: every figure is computed in TypeScript tools against the ranking ledger; the model only chooses which tool to call
- Keep in mindProject in progress; tool choice is benchmarked across LLMs and answer wording is still generated