Inspiration
First of all, ** when watching professional games or reviewing data, we find a long-standing problem: ** There is a lot of data, but there is almost no system that really ** "talks like a coach". The existing analysis tools either only give raw statistics (K/D, economic curve) or directly generate generalised AI summaries, **cannot answer the questions that coaches are really concerned about, such as:
- Is this a high-risk game? ** - ** What exactly is the economic problem? ** - ** If a key player dies in the early stage, what kind of chain reaction will occur in the whole game? ** We gradually realised that the problem is not "whether there is AI", > but whether the analysis system really understands the complete process of > **"question → judgement → evidence → conclusion". Drift Coach was born from this - an analysis system with coaching thinking as the core, not model output as the core. --- ## What it does
Then, Drift Coach is a natural language-driven e-sports game analysis and replay system. ** Users can ask questions directly in natural language, such as: - "Where is the economic management problem of this game?" - "Please give me the review agenda of this game" - "Is this a high-risk game?" - "Please summarise the key lessons of this game" The system will be in order Complete: 1. **Understand the intention of the problem (judgement / analysis / review / summary) 2. Excavate evidence related to the problem in the data 3. Generate a structured conclusion (fact → impact → suggestion) 4. ** Clearly mark the low confidence when the evidence is insufficient, instead of fabricating the answer** The final output is: > The analytical text that can be directly used for the coaching meeting > Instead of the original output of logs, tables, or models. ---
How we built it
Then, we split the system into multiple levels with clear responsibilities, instead of a "black box AI": 1. Intent Parsing Map natural language problems to analytical intentions (such as MATCH_REVIEW, ECONOMIC_ANALYSIS, RISK_JUDGMENT). 2. Fact Mining Extract Verifiable facts from the game event data (round, economy, target control, risk sequence, etc.). 3. Derived Findings does not directly output facts, but generates problem-level conclusion unit as what the system really "knows". 4. Question-level State Each question has an independent state to prevent the answers to different questions from contaminating each other. Five. Narrative Synthesis generates narrative text that can be read by the coach only based on Derived Findings, rather than mining data again. 6. Frontend Safety Layer Defence rendering, abnormal bottom, security mode, to ensure that the Demo can be displayed stably. The whole system always emphasises one sentence:
Judgement precedes narrative, and evidence precedes conclusion --- ## Challenges we ran into Then, we encountered some key challenges: ### 1. Different questions get the same answer. At the beginning, review questions and economic questions will output almost the same content. → We introduce Question Scope Reducer and Derived Findings Pool to force each question to see only the conclusions related to themselves. ### 2. Should the system "speak" when the evidence is insufficient? Completely refusing to answer will affect the experience, and forced answering will be misleading. → Final choice: low confidence, clear labelling, unchanged structure. ### three. Too much data causes the system to crash. A query may trigger thousands of events. → Introduce factual whitelist, content trunction, narrative safety valve. ### 4. Front-end rendering stability, complex structure + long text is very easy to cause front-end crash. → Realise Error Boundary, security mode and original content back.
Accomplishments that we're proud of
What I want to say most is: - ✅ The system can ** directly answer coaching questions in natural language ** - ✅ Different questions produce ** different and non-polluting answers ** - ✅ Judgemental questions (such as "high risk") coexist with analytical questions - ✅ Clear distinction ** Sufficient evidence / insufficient evidence / low confidence assumption** - ✅ During the Demo process, the system will not collapse due to abnormal data. The most important thing is: > ** The system will honestly say that it does not know when it "does not know". **---
What we learned
In fact, what we have learned is not "how to make AI talk better", but: - The most difficult thing for AI analysis system is not to generate text, but to control when not to generate it - "Intelligent" does not mean "talk more", but to say the right words within the right range - Structured reasoning is more important than the scale of the model - A useable system must be responsible for Demo experience and understanding cost
What's next for drift coach
Next, we hope to let Drift Coach really grow: - 🔹 Let the analysis results feed back the follow-up problems (continuous learning instead of pollution) - 🔹 Introduce Comparative analysis across competitions and cross-seasons - 🔹 Support more games and more Fine-grained tactical hierarchy - 🔹 Make Drift Coach a long-term analytical partner to accompany coaching decision-making Our goal is not: > "Be a smarter AI", but: > ** Make a more reliable coaching system. **
Log in or sign up for Devpost to join the conversation.