Project 02 of 06·May — Jun 2025

TechTree

The ultimate mobile-games market-intelligence platform — understanding every game in depth, triangulating data that today lives in silos across GameRefinery, SensorTower, and App Annie.

Product Mobile Games B2B Market Intelligence Computer Vision Founder INACTIVE

01Artifacts

▶ Demo & artifacts · coming soon

Mobile-game taxonomy

7 domains · 35 categories

Coming soon

Gameplay agent

YOLO + AFK Arena autonomous run

Coming soon

Chatbot analyst

Streamlit + Agents SDK demo

Coming soon

Pipeline architecture

Per-frame → whole-video rewrite

Coming soon

02At a glance

Role

FounderSolo

Timeline

~4 weeksMay 23 — Jun 17, 2025

Codebase

~39k LOC4 repos · 124 commits

Stack

Gemini · GPT-4pgvector · ChromaDB · Cohere

03The work

Mobile-game product teams rely on fragmented tools today — GameRefinery for feature taxonomies (curated by human analysts), SensorTower for revenue, App Annie for rankings — and none of them actually watch the games. TechTree’s vision was the ultimate platform that understands every mobile game deeply, triangulating the signals those tools fragment.

The product was built as three coordinated systems. A live gameplay capture agent instruments a running mobile game on an Android device or emulator, records every screen and touch interaction as it happens, and auto-annotates each screen against a hand-built 7-domain / 35-category mechanic taxonomy (core loop, monetization, live ops, progression, competitive, UX, analytics). A video analysis pipeline does the same job at scale from public YouTube gameplay, ingesting whole videos, deduplicating frames, and populating a PostgreSQL + pgvector feature database. An AI chatbot analyst sits on top, answering natural-language questions like “show me all the gacha systems in match-3 games” by combining semantic search with SQL joins across the feature, screenshot, and taxonomy tables — and returning explanations with the annotated screenshots inline.

04Skills

Market intelligence Computer visionYOLO Data modeling Product taxonomy design7 × 35 Conversational AI Agent-based systems Semantic search Vector databases Prompt engineering LLM evaluation Mobile automationAndroid · adb Batch orchestration900/min

05Notable accomplishments

Autonomous gameplay agent

An agent that plays a mobile game end-to-end on an Android emulator and maps its features live — no human required. Trained a YOLO computer-vision model on actual AFK Arena gameplay UI (buttons, menus, state elements), hand-labeled, so the agent could recognize the interface and navigate the game autonomously. Every screen and action logged in real time; each screen auto-annotated against the 7-domain mobile-game taxonomy via a GPT-4 vision pass. Solves the hardest problem in mobile-games intelligence: getting structured feature data out of a live game without manual analyst labor. Built on uiautomator2 + adb for device control, YOLO for UI recognition, GPT-4 vision for screen canonicalization. GameMarketer_HybridProcessingFlow/

Video analysis and feature extraction pipeline

Ingests public YouTube gameplay videos and lands structured feature data in a queryable database. End-to-end: yt-dlp downloads, FFmpeg + perceptual hashing deduplicates frames (91% volume reduction), Gemini 2.5 Pro with context caching annotates the whole video against the taxonomy in a single pass, and results land in PostgreSQL with pgvector embeddings. The v1→v2 rewrite inverted the architecture — from per-frame annotation to whole-video reasoning — and saved two API calls per feature. UnifiedFlow/holistic_annotation_flow_v2.py

AI chatbot analyst

A natural-language interface over the feature database. A user asks “show me all the gacha systems in match-3 games” and the agent answers with feature explanations and the actual annotated screenshots inline. Under the hood: a Streamlit app with an OpenAI Agents SDK agent that translates the question into semantic search over ChromaDB, a Cohere rerank pass, and SQL joins across features_game, screenshot_feature_xref, and taxonomy. A Pydantic AgentResponse(user_response, developer_note) structure gives every turn a hidden meta-channel that surfaces the agent’s own prompt-quality concerns back to the PM. GameMarketer_Frontend/utils/agent_config.py