DiceTree

Turn scattered datainto AI that works

DiceTree designs, builds and runs generative AI and RAG systems, data platforms and analytics, and AX PoCs. We built a standard generative AI development platform for an enterprise client and an automated VOC response system for an airline.

Ask about an AX PoC See projects
Databases, documents, audio and images are collected, parsed and refined, gathered in a unified store, and used for RAG search, VOC auto-replies, natural-language data search and agent workflows Collect Parse Refine Store Use Databases Documents, PDF Audio, video Images Parse, recognize VLM, OCR, Whisper Clean, embed Chunking, lineage Unified store Vector, graph, BigQuery RAG search, chatbot VOC auto-reply Natural-language search Agent workflows Pipeline from data collection to AI use DB Docs Audio Images Parse, recognize VLM, OCR, Whisper Clean, embed Chunking, lineage Unified store Vector, graph, BigQuery RAG chatbot VOC auto-reply Data search Agent workflows
How a company's databases, documents, audio and images become usable by AI agents. We designed this structure in an enterprise RAG data store consulting engagement. Since 2023

What we do well

We build everything ourselves, from planning to deployment and operations. Each area below has projects we delivered or still run.

  • Generative AI and RAG systems

    Key projects Generative AI development platform, VOC auto-reply, RootBox, rootPDF

    RAG that answers from your documents and data, and agent workflows that switch between LLMs, built all the way to production.

    • Document ingestion and indexing pipelines with an admin console
    • Automatic draft replies to customer inquiries
    • Node-based agent workflow editor
    • Tables and formulas extracted from Korean PDFs with a VLM
    • FastAPI
    • React
    • TypeScript
    • LangChain
    • ChromaDB
    • pgvector
    • Docker
    • GitLab CI

  • Data platforms and analytics

    Key projects Data Governance Suite, NPS category redesign, HomeScore

    We manage metadata, lineage and access on BigQuery data platforms, and analyze the data to build classification criteria and metrics.

    • Data flow tracing with our own SQL parser
    • Automated metadata collection and table access control
    • Classification criteria checked against a gold set
    • Spatial scoring and weighting models
    • Python
    • BigQuery
    • GCP
    • PostgreSQL
    • PostGIS
    • Streamlit
    • PyTorch

  • AX consulting and PoC

    Key projects Enterprise RAG data store design, rootPDF, decision model fine-tuning

    We decide which work AI should take on, build it small, prove it with numbers, then hand it over to operations. We also cover architecture, effort estimates and executive reports.

    • Enterprise data integration architecture
    • Evaluation tooling with token and cost measurement
    • Training and evaluating task-specific decision models
    • Handover documents and production rollout

  • Apps and services

    Key projects HomeWallet, MiniCaddy, Daily CSAT, Sinamu

    We plan, build, launch and run Android, Galaxy Watch (Wear OS) and web services ourselves.

    • Store releases and signed builds
    • Phone-to-watch sync over Bluetooth
    • Map and location-based services
    • Web apps where an LLM creates new content every day
    • Kotlin
    • Jetpack Compose
    • Wear OS
    • Next.js
    • React
    • Supabase
    • Cloudflare

  • AI video and voice content

    Key projects Caba & Friends, Shorts Gen, HAVIS

    We use generative video and speech synthesis to make animation, short-form video and AI assistants with a real voice.

    • Animation made with an AI video pipeline
    • Automatic Shorts editing, narration and upload
    • Speech synthesis from a cloned voice
    • Lip-synced avatars
    • ComfyUI
    • Veo
    • Whisper
    • GPT-SoVITS
    • FFmpeg
    • Unity

How we run an AX PoC

AI adoption costs less to get wrong when you test small and decide on numbers. We split the work into five steps, and we have done each one on a real project.

  1. Assess work and data

    Pick the tasks where AI can help, and find out where the usable data lives and in what shape.

    ExampleData Seeker, natural-language search over internal data

  2. Design the architecture

    Define the data flow and storage from ingestion to the knowledge layer, and estimate the effort.

    ExampleEnterprise RAG data store design, CTO report delivered

  3. Build the PoC

    Build a working system with only the core features and run it on real data.

    ExamplerootPDF, a VLM-based Korean PDF parser

  4. Measure

    Score accuracy against a gold set, and work out token costs and the serving setup.

    ExampleNPS gold set validation, evalkit evaluation tool

  5. Move to production

    Automate deployment with Docker and CI, and hand over the documentation.

    ExampleGenerative AI platform deployed for several clients

Have a task in mind? We start by looking at your data together.

Ask about an AX PoC

Data analytics work

We trace where data comes from and where it goes, set the criteria, and automate decisions with models.

  • Table-level data lineage

    Our own SQL parser (PLY) reads BigQuery scripts and execution logs to trace which tables are built from which.

    BigQueryPLYPython

  • Metadata and access governance

    BigQuery metadata is collected into the in-house metadata system, and table access is managed through Data Catalog.

    GCPData CatalogFastAPI

  • VOC classification redesign

    We redefined an airline's NPS categories, validated the reference data (gold set) with GenAI, and wrote classification rules R1–R4.

    GenAIGold setRAG

  • Support ticket decision model

    We fine-tuned and evaluated a dedicated decision model on 10,000 Korean support tickets.

    PyTorchModernBERT

  • Livability scoring

    Transport, healthcare, education, shopping and nature within 500 m and 1 km are scored, with weights for different lifestyles.

    PostGISSupabaseNext.js

  • Natural-language data search

    RAG search that finds internal data from questions like “Where is last month's sales table?”

    LangChainpgvectorStreamlit

Projects

From AI services and enterprise data platforms to mobile apps and game and content IP, these are projects we took from planning to deployment and operations. Select a project to see the details.

Generative AI development platform

AI agents and a standard generative AI development environment for an enterprise client 2025.02 – 2025.12
  • RAG pipeline for document ingestion and indexing, with an admin web console
  • Standalone and general-purpose editions tailored to several clients
  • Build and deployment automated with Docker and GitLab CI, SSL enabled
  • React 18
  • TypeScript
  • Vite
  • MUI
  • Zustand
  • FastAPI
  • Docker
  • GitLab CI

Customer VOC classification and reply

A system that analyzes an airline's customer VOC with RAG and drafts replies automatically 2025.04 – 2025.12
  • VOC indexing and workflow pipeline
  • Automatic reply generation, several production releases
  • Follow-up: NPS category redesign with gold set validation, classification rules R1–R4 and handover documents (2026.05 – 06)
  • FastAPI
  • RAG
  • Docker
  • GitLab CI
  • Claude

RootBox

Chat, RAG and a node-based agent editor across multiple LLM providers 2026.02 – 2026.06
  • Choice of LLM providers, streaming chat, image and PDF input
  • RAG with ChromaDB and a canvas for wiring nodes into workflows
  • Structured PDF extraction, Swagger auth, API key encryption (Fernet)
  • Earlier version built on AWS Bedrock, Vertex AI and OpenSearch
  • FastAPI
  • SQLite
  • ChromaDB
  • sentence-transformers
  • Vanilla JS

Enterprise RAG data store design

Architecture and consulting to make a company's databases, documents, audio and images usable by AI agents 2026.01
  • Five-stage pipeline: ingestion, parsing, processing, unified storage, knowledge layer
  • Hybrid vector and graph design, AWS-based multi-cloud, data governance policy
  • CTO report and effort estimate delivered

Data Governance Suite

Tools for running a data platform: BigQuery metadata, lineage and access 2024
  • Metadata collection: BigQuery metadata converted to the in-house metadata format
  • Data lineage: table-to-table flows traced with our own SQL parser (PLY)
  • Access control: BigQuery table permissions linked with Data Catalog
  • Dynamic API Server: register and run APIs at runtime, key management, async jobs
  • Data Seeker: RAG search over internal data in natural language
  • Python
  • BigQuery
  • GCP
  • FastAPI
  • Streamlit
  • LangChain
  • PostgreSQL

rootPDF

A VLM-based Korean PDF parser PoC that reads each page as an image 2026.06
  • Avoids broken Korean text by using a VLM to read layout and reading order
  • Tables to Markdown, formulas to LaTeX, figures to descriptions
  • Token and cost measurement, serving simulator, evaluation tool (evalkit)
  • Python
  • PyMuPDF
  • Claude Vision

Haxcel

An Excel add-in that pulls server data straight into the sheet to automate work 2023 – 2025
  • C# VSTO client with a FastAPI server
  • Web add-in version prototyped with Office JS
  • C# VSTO
  • FastAPI
  • SQLite
  • Office JS

VRoid MCP

An MCP server that lets Claude Code drive VRoid Studio, which has no public API 2026.10
  • Reverse-engineered .vroid format to edit sliders and colors
  • .vrm editing and texture swaps, Blender preview renders and batch jobs
  • Live control of VRoid Studio through on-screen text recognition (OCR)
  • Python
  • MCP
  • Blender

Shorts Gen

YouTube Shorts automation from source search to editing, narration and upload 2026.02
  • Collection with yt-dlp, transcription with faster-whisper, highlight picks by an LLM
  • Korean narration by speech synthesis, 9:16 captioned clips with FFmpeg
  • Automatic YouTube upload and scheduler
  • Python
  • FFmpeg
  • faster-whisper
  • GPT-SoVITS

HomeWallet Live

An Android budget app for managing card spending with your family 2025.10 – 2026.09
  • Card transactions, search across months, shared users
  • Per-user data isolation with Supabase row-level security, signed release builds (v1.14)
  • Kotlin
  • Android
  • Supabase

MiniCaddy

A Galaxy Watch golf caddy that shows distance to the pin and a hole map on your wrist 2026.06 – 2026.09
  • Distance to the pin, hole number and par, scorecard, hole minimap
  • Phone companion app sends server data to the watch over Bluetooth
  • Pipeline and editor that draw hole maps from satellite imagery
  • Kotlin
  • Compose for Wear OS
  • Python
  • OpenStreetMap

Face Quiz Ready to launch

A mobile puzzle and quiz game played with your own face 2026.07 – 2026.08
  • Face part masking, blink controls, shareable result cards
  • Three game modes, store listing materials ready
  • Kotlin
  • Android
  • Face detection

PathCare / WRU

Location tracking for children and older adults, and location sharing in the family 2026.01
  • Real-time GPS tracking, FCM emergency alerts
  • App for the protected person, web app for guardians, admin dashboard, landing page
  • Kotlin
  • Jetpack Compose
  • Flask
  • Firebase
  • Next.js

Daily CSAT

A five-minute daily study app where AI writes new Korean SAT-style questions every day 2026.02
  • An LLM writes new questions for each subject at 5 a.m. every day
  • Rankings, wrong-answer notes, study analytics, community, Google sign-in
  • Docker setup and an Android app
  • React
  • Vite
  • Tailwind
  • FastAPI
  • SQLAlchemy
  • Firebase Auth

HomeScore

Drop a pin on the map to score nearby amenities on a radar chart 2026.08 – 2026.09
  • Transport, healthcare, education, shopping and nature scored within 500 m and 1 km
  • Nationwide data, rankings, weights by lifestyle
  • Next.js 16
  • React 19
  • Tailwind v4
  • Kakao Map
  • PostGIS
  • Vitest

Sinamu (poem_tree) Live

A poet's archive with an AI recitation theater, where poems are read over AI-made scenes 2025.05 – 2026.09
  • 81 scanned pages of poetry collections restored, upscaled and digitized with OCR
  • Recitation theater: audio and captions aligned with Whisper, scenes planned by an LLM, AI-generated images
  • Node.js
  • Cloudflare Pages
  • Workers AI

Startem

Commerce curation of items worn by celebrities 2023.08 – 2023.11
  • Categories, filters and product pages
  • Product detection in images with YOLO
  • Flask
  • React
  • SQLite
  • YOLO

KeeTok

A social service designed to be safe for children 2024.01 – 2024.07
  • Web service and Android app released
  • FastAPI
  • Nginx
  • Android

HAVIS

An AI assistant with your own voice, using speech recognition, an LLM and speech synthesis 2025.12 – 2026.02
  • Speech synthesis from a cloned voice
  • Lip-synced avatar, Android app
  • Whisper
  • GPT-SoVITS
  • LLM
  • Wav2Lip
  • LivePortrait

Caba & Friends

A capybara character IP spanning children's animation and a web game 2025.12 – 2026.02
  • 101-second opening and theme song made with an AI video pipeline (ComfyUI, Veo)
  • Kapibara Survival: a web action game MVP with 5 weapons, 4 enemy types, 11 waves and a quiz mode
  • ComfyUI
  • Veo
  • Phaser 3
  • TypeScript

SaintKeeper

A playable 3D action RPG prototype about protecting a saint 2023
  • Combat and damage, skill tree and equip UI, database-driven skill data
  • Multiplayer experiment with Mirror, four VRoid characters
  • Unity
  • C#
  • Mirror
  • MySQL

Golf Princess In development

A 2D raising sim where a dad who knows nothing about golf turns his daughter into a major champion 2026.09 –
  • Training schedules, caddying on tournament day (club and strategy picks), choosing a coach
  • TypeScript
  • Vite

DiceTree Kids

Interactive content for children: kids' drums, a cappella, whack-a-mole and more 2023 – 2024
  • Served from the company website
  • Automatic deployment to our own server with GitHub Actions

Research and experiments

Topics where we tested the technology ourselves before it became a product.

TopicWhat it doesStack
AI TradingReal-time market data, ML forecasts, automated trading, dashboard, Telegram control (long-running)Kafka, Flask, Dash, ML
MiniTrackBall trajectory tracking in golf swing videos (YOLO, stabilization)FastAPI, OpenCV, YOLO, React Native
SwingDaddyGolf swing posture analysis with pose estimationMediaPipe, OpenCV, FastAPI
Decision modelFine-tuning and evaluating a dedicated decision model (10,000 Korean support tickets)PyTorch, ModernBERT
CubeXResearch on a compression algorithm that places data in 3D spacePython, NumPy
FoodinatorFood recommendations from answers to a few questionsFlask, KoGPT2
AI Code EditorPrototype of a Cursor-style AI code editorElectron, Monaco

We take on AX PoCs, data platforms and app development. Tell us about the work you have in mind.

admin@dicetree.com
github.com/dicetree-admin www.dicetree.com