TARCRAFT II AI COACH — PROJECT PROPOSAL

STARCRAFT II AI COACH — PROJECT PROPOSAL
=========================================

ONE-LINE PITCH
--------------
A desktop coaching app that watches your StarCraft II games via screen
capture (no game API, no injection), coaches you in real time in the
voice and style of a partnered streamer, reviews your games afterward,
and connects you to that streamer for paid 1-on-1 sessions with the
prep already done.


THE PROBLEM
-----------
- Most ladder players plateau on macro fundamentals: supply blocks,
  late expansions, floating resources, no creep spread, idle workers.
- Human coaching is the fix but it is expensive, and half of a paid
  hour is the coach watching replays to find out what is wrong.
- Build order guides are static. They do not know what you actually
  did in your last 40 games.
- There is no supported way to read live ladder game state through the
  official API. The only path to a live coach is machine vision.


THE PRODUCT
-----------
1. LIVE COACH (runs locally, real time)
   - Screen capture of the HUD and minimap at fixed positions.
   - Small in-house neural nets read minerals, gas, supply, timer,
     worker count, production tab, creep coverage (minimap pixels).
   - Rule-based cue engine fires short audio barks in the streamer's
     voice: "supply", "inject", "gas", "you're floating", "army home".
   - Build order tracking keyed on supply count, with hard steps
     (always do X at Y supply) and conditional steps (hold unless Z
     seen) to encode "when to wait".
   - Small overlay showing next two build steps and progress marker.
   - Cue escalation and fight detection to avoid nagging.

2. POST-GAME REVIEW
   - Full timeline from the vision layer plus the replay file
     (replay parsing is fully supported and contains complete unit
     data, so review can cover what live vision cannot).
   - Local LLM (7-8B, fine-tuned per streamer) writes a review in the
     streamer's voice using their coaching rubric.
   - Tracks trends across games: recurring mistakes, what is fixed,
     what is not.

3. STREAMER PERSONAS
   - Each partnered streamer ships as a versioned package: decision
     model, coaching rubric, recorded voice lines, build library.
   - Persona decisions come from the streamer, not from generic
     strategy advice.

4. DIRECT-TO-STREAMER COACHING
   - "Book a session" button inside the app.
   - Before the session, the AI hands the streamer a brief: skill
     trend, top three recurring problems ranked by cost, what is
     already fine, the three replays that show it best with
     timestamps, and what the AI persona has been telling the player.
   - Streamer walks in prepped. The whole hour is teaching.
   - After the session, streamer marks what was covered and the AI
     persona adjusts emphasis for that player.
   - Player sees the same brief. Nothing about them goes to a third
     party unseen.


HOW A PERSONA IS BUILT (WITH THE STREAMER, NOT FROM SCRAPING)
--------------------------------------------------------------
- HOW THEY DECIDE: parse the streamer's own replay archive into
  state -> action pairs. Train a small decision model and/or mine
  statistical rules ("at 30 supply with no forward pylon seen, third
  base 94% of the time"). Combine with interview-style rubric
  extraction: what do you look at first, what is fatal vs cosmetic,
  what is your fix for each.
- HOW THEY TALK: LoRA fine-tune of a small local model on transcripts
  of their coaching/commentary content, filtered to coaching-relevant
  segments. Rubric in the prompt, fine-tune for voice. Kept separate
  so either can be updated alone.
- HOW THEY SOUND: recorded lines for frequent cues; optional consented
  voice clone from clean studio audio for long-tail narration.
- QUALITY LOOP: streamer reviews generated output in batches and marks
  what they would not say. That correction data is the primary
  training signal over time.
- The same replay pipeline run on the PLAYER's games shows what the
  player actually does. The gap between streamer model and player
  model, expressed in the streamer's voice, is the coaching product.


ARCHITECTURE / COST CONTROL
---------------------------
- Everything real-time runs on the player's machine: vision, cues,
  TTS, rules. SC2 leaves the GPU mostly idle, so this is free.
- Post-game review runs locally by default. Optional cloud tier with a
  stronger model, priced to cover its own cost.
- Cloud footprint is minimal: accounts, payments, bookings, streamer
  payouts, build-order library, auto-update.
- Variable cost per user is near zero. Margins do not shrink as it
  grows.
- Minimum spec plus a lite mode (rule-based review only) for weaker
  hardware.
- Do not build accounts, payments, or scheduling in-house. Stripe
  Connect plus an off-the-shelf booking tool for year one.


BUSINESS MODEL
--------------
- The company owns the app, vision engine, coaching logic, platform.
  Streamers are licensing partners, not sellers.
- SUBSCRIPTION: ~$5-10/month for the AI coach. When a subscriber uses a
  streamer's persona, that streamer receives a revenue share
  (~20-40% of that subscription). Generic personas are 100% platform.
- BOOKING FEE: streamer sets their 1-on-1 price; platform takes
  ~15-25%.
- Pitch to streamers leads with pre-briefed sessions ("I double your
  effective coaching rate and send you players who already know what
  they need"), not the revenue share.
- Rough shape: 1,000 subs at $8 = ~$8k/mo gross, ~$4-5k net after
  shares and costs. 50 sessions/mo at $50 with 20% = ~$500. Niche,
  real, grows if it works. Engine ports to other games later.


LEGAL / TRUST
-------------
- No streamer likeness, voice, or content used without a signed
  agreement. Original personas only until partners are onboard.
- Training rights explicit in writing: what content may be trained
  on, model exclusive to the app, persona retired and model deleted
  if the streamer exits, streamer approval over new persona behavior
  before ship.
- Verify current Blizzard EULA language on passive screen capture and
  overlays before launch. No input injection, no memory reading.
- Player gameplay stays on the player's machine unless they opt in.


ROADMAP
-------
PHASE 0 (weekend): capture HUD at 1080p, read minerals/gas/supply/
  time, fire a TTS cue on float or supply block. Does it feel good
  in a real game? Everything depends on this answer.
PHASE 1: full HUD + minimap reading, build order tracker with
  hard/conditional steps, one generic persona, overlay, replay-based
  post-game review with a local model.
PHASE 2: form the entity, sign one or two mid-sized streamers, build
  the first real persona (replay mining + rubric + recordings + LoRA),
  charge from day one.
PHASE 3: booking + AI pre-brief pipeline, streamer dashboard, payout
  automation, post-session feedback loop.
PHASE 4: more streamers, more races/matchups, optional cloud review
  tier, evaluate second game.

SUCCESS CHECK: 200 paying users within a few months of Phase 2. If
yes, push. If no, you still have a working vision engine and a fun
tool.


KNOWN LIMITS
------------
- Live vision is strong on macro, weak on micro. Do not promise micro
  coaching live; replay review covers it after the fact.
- Local-only distribution weakens subscription enforcement. The
  streamer content and 1-on-1 bookings are the parts that genuinely
  require the platform.
- Small market. This is a solid side business, not a rocket.