Songstark Logo

WHAT WE DO

Engine to product, direct.

AI engine builder and full-stack product team. We fuse AI agents and automation into every step — planning, design, development, operations — and go faster and deeper at the same quality bar.

Three lanes.

One lane for building the engine itself. One lane for carrying the product to the finish. Mobile & edge AI runs alongside as a third specialty.

  1. 01

    We build the engine ourselves

    Not an agency-style delivery. We design, build, and patent voice-analysis, LLM-report, and RAG pipelines in-house. 5 voice-analysis patents filed (67 claims · 34 figures), 10 total patents filed across the company.

    Primary stack

    • Voice AI
    • 3-Pass LLM
    • pgvector
    • Python · PyTorch
    • LLM Ops

  2. 02

    We run the product to the finish

    Design, web, mobile, admin, payments, deployment — closed inside a single team. AI isn’t required on day one, the quality still holds. Public-sector requirements (KWCAG AA · ISMS-P) ship by default.

    Primary stack

    • Next.js 16
    • React 19
    • Supabase
    • Cloud Run
    • Expo

  3. 03

    Mobile, same team

    Push trained AI models out to mobile and edge devices. On-device inference, sensor/IoT integration, and low-bandwidth / offline handling all included.

    Primary stack

    • Expo
    • Kotlin · Android
    • Swift · iOS
    • On-device inference
    • BLE · IoT

Case studies

Things we’ve made that are still running today.

From our own SpeechMap SaaS to products running with KAIST, Gangwon Office of Education, BAND Foundation, Dururbarun Co-op, and the Self-Reliance Cooperative Federation — every case here is live.

AI voice-analysis engine · SaaS · 2025–2026

SpeechMap · in-house SaaS

Problem
The domain needed an analysis engine that reads child language development through Jitter/Shimmer/HNR acoustic features and MFA phoneme accuracy.
Impact
4-Path pipeline · 12+ clinical indicators · 5 patents (67 claims · 34 figures) · Cloud Run 4GB/2vCPU in production

Stack

  • Next.js 16
  • Python · Parselmouth
  • Google STT
  • MFA · GPT-4o-mini

Research platform · core language-development screening · 2024–2026

Gangwon Office of Education · speech-language pathology research team

Problem
Age-banded K-DST screening and a standardized checklist had to run and auto-score online, and the path through in-depth testing and therapy support had to live in one platform.
Impact
111 age-banded items · 3 roles (parent · clinician · researcher) · anonymization + institutional RLS · in-depth testing and therapy support on the same platform · 111 PRs on language-screening-platform-2025, actively maintained

Stack

  • Next.js
  • Supabase RLS
  • TypeScript
  • Research Ethics

Social-entrepreneur capability diagnostic · digital twin · 2025–2026

BAND Foundation

Problem
Social entrepreneurs needed a standardized capability diagnostic, and their investors needed a digital-twin platform to track and manage the portfolio.
Impact
Entrepreneur diagnostic · investor portfolio management · digital twin · Expo mobile companion shipped by the same team

Stack

  • Next.js
  • Supabase
  • Expo
  • GPT-4o-mini

Biofoundry education platform · AI tutor · 2026 · Ongoing

KAIST

Problem
Graduate students and researchers needed a step-through synthetic-biology LMS where an AI tutor answers in context.
Impact
Contracted by KAIST · in progress · research-grade LMS · embedded AI tutor

Stack

  • Next.js 16
  • Supabase
  • pgvector RAG
  • TypeScript

Official-site renewal · interest-support program · 2025–2026

Woorihamkke · Korea Self-Reliance Cooperative Federation

Problem
Member onboarding, mutual-aid programs, and interest-support cases had to consolidate into one renewed federation website.
Impact
6-step interest-support workflow · 6 mutual-aid programs · admin CMS · legacy notice/library migration finished

Stack

  • Next.js 16
  • Supabase RLS
  • TypeScript
  • Tailwind

B2B merchandise consulting platform · 2026

Gong Gong Gong Gan · Withgoods

Problem
Corporate-merch planning, production, and delivery had to flow through a single B2B consulting site end-to-end.
Impact
2-step wizard lead form · LLM-pipeline diagnostic · case-gallery CMS · admin dashboard · multi-language marketing pages

Stack

  • Next.js 16
  • Supabase
  • TypeScript
  • Tailwind

Mobile puzzle RPG · iOS · Android · 2026

Songstark · in-house IP

Problem
A medieval-fantasy story puzzle RPG had to ship from in-house concept through level design and localization on a single codebase.
Impact
8 chapters · 410 puzzle levels · season pass + daily challenge · 6-locale release · Expo single codebase

Stack

  • Expo SDK 55
  • React Native
  • TypeScript
  • i18n · 6 locales

Carbon-credit MRV · web + Android · 2026

Kijamitable · Kijanify

Problem
A single team had to run carbon MRV end-to-end — from Rwandan field measurement to credit issuance — across web and Android.
Impact
Earth Engine integration · field-ready Kotlin Android · web dashboard · Seoul–Kigali one-team delivery

Stack

  • Next.js
  • Kotlin · Android
  • Google Earth Engine
  • TypeScript

HOW WE WORK

AI moves us fast; humans set the bar.

Four steps from kickoff to ongoing operations. AI agents accelerate each step; humans lock the quality bar.

  1. STEP 01

    Understand

    AI research reads the market, competitors, and user context fast.

    • Kickoff workshop
    • LLM competitive analysis
    • IA auto-draft
  2. STEP 02

    Design

    AI prompts align structure and flow quickly.

    • Wireframes
    • Data modeling
    • Design system
  3. STEP 03

    Build

    Code agents and engineers ship together.

    • Copilot workflow
    • CMS · admin
    • Automated QA · perf
  4. STEP 04

    Operate

    AI monitoring makes post-launch faster than launch.

    • Anomaly detection AI
    • Auto improvement sprints
    • Outcome reports

TECH

From AI to infra, full-stack.

LLMs, vector search, and agent workflows — baked into the entire production pipeline.

Frontend

  • Next.js · Nuxt
  • React · Vue
  • TypeScript
  • TailwindCSS
  • Framer Motion

Backend

  • Node.js · NestJS
  • Spring Boot
  • Python · FastAPI
  • GraphQL · REST
  • WebSocket

AI & ML

  • OpenAI · Claude
  • LangChain · LlamaIndex
  • Vercel AI SDK
  • pgvector · Pinecone
  • Hugging Face

Infra & Ops

  • AWS · GCP
  • Vercel · Cloudflare
  • Docker · K8s
  • GitHub Actions
  • Datadog · Sentry

REFERENCES

Places that trusted us with the work.

Public institutions, universities, research labs, cooperatives, and global startups — references accumulated across domains.

Public · Research · Education

KAIST, Gangwon, BAND Foundation, Dururbarun Co-op — domains where trust is non-negotiable.

  • Gender-equity education platform

    KIGEPE · attendance · high-volume processing · AI chatbot

    2026

  • SMIT entrepreneurship practicum platform

    SMIT · NABCD 31-item AI review · professor dashboard

    2026

  • Eunpyeong social-solidarity economy hub

    Eunpyeong-gu · website renewal · admin console

    2026

  • Onju Literature official site

    Literary society · writing-contest workflow

    2025–2026

AI · SaaS products

From our own engines to healthcare SaaS and AI studios — built and operated in-house.

  • Medifood Platform · Plate Clinic

    Clinic brand · multi-tenant report · NDA

    2026

  • Veltis AI studio

    Script → slides → avatar video automation

    2026

  • history-book · AI interactive fiction

    LLM streaming gamebook · image generation

    2026

  • a11y-checker · accessibility diagnostic SaaS

    axe + Gemini fix suggestions · KWCAG

    2026

  • Marea Holdings sales automation

    Next.js 16 RSC · Cloud Run · AI email · auto quoting

    2026

  • GPT self-compassion writing research

    IRB-compliant research survey · PANAS-SF · paid honorarium

    2025

B2B · Commerce

Brand sites, e-commerce, and operational automation — carrying company growth.

  • K-mart IDP commerce

    Django + React · admin WYSIWYG

    2024

  • Minacaps – specialty hat store

    ko/en/ja · next-intl · OG/SEO

    2025

  • Stark Academy

    In-house education platform · Tiptap · Supabase

    2025–2026

How it’s different, on one page

What working with this team actually looks like.

The axes that actually decide RFPs, PoCs, and proposals. Each row surfaces the difference between a firm that adopted AI and a team that has built AI.

  • AI engine

    Ordinary dev shop
    Thin wrapper over OpenAI · Azure APIs
    This team
    Our own patented engine SpeechMap + LLM · RAG design

  • Academic grounding

    Ordinary dev shop
    No professor involvement, or one advisory name
    This team
    9 in-house patents filed · paper-oriented R&D track

  • Product vs. services

    Ordinary dev shop
    Services-only delivery, no in-house AI product
    This team
    We run our own AI product and take client engineering work in parallel

  • Maintenance

    Ordinary dev shop
    Separate contract, flat monthly retainer
    This team
    Live maintenance repo pointers (e.g. language-screening-platform-2025)

  • Communication

    Ordinary dev shop
    Sales PM → in-house planner → subcontracted dev, 3–4 hops
    This team
    Direct line to the CEO and the core engineer

  • Delivery verification

    Ordinary dev shop
    Post-delivery acceptance testing
    This team
    Vitest · Playwright CI + Sentry runtime monitoring

  • Accessibility & privacy

    Ordinary dev shop
    “KWCAG AA compliant” declared without proof
    This team
    KWCAG AA · PIPA · ISMS-P by default · ready-to-ship Supabase RLS policies

Want to talk?

Send us the problem you’re trying to solve. No sales PM in the middle — the CEO or a core engineer replies directly within one business day.

Send us your projectRequest a SpeechMap demo