Skip to content
Youngtak Jo
Contact

Youngtak Jo

CV

I am a founder, product owner, and full-stack builder focused on turning complex, unstructured real-world problems into scalable, production-grade software platforms.

Download PDF

Experience

Seoul

Wrtn Technologies Current

Agent Developer, Coding Agent Team

I lead ontology-driven agent development on the Coding Agent team. I continue to develop Consilience, which I had been building alone before joining, and apply the knowledge-graph, retrieval and agent design validated there to the team's coding agent.

  • Lead ontology-driven AI agent development on the Coding Agent team.
  • Continue to develop Consilience and apply its knowledge-graph code intelligence to the team's product.
  • (to be filled in — specific outcomes and metrics at Wrtn)

TypeScript · Rust · LLM Agents · RDF/SPARQL · tree-sitter

Projects

Consilience

Seoul

Independent — solo product Ongoing

Solo founder and engineer — Consilience

I do all of it alone: the research and design, the Rust and TypeScript implementation, the public benchmarks and their write-ups, a signed and notarised desktop release with auto-update, the cloud backend and subscription billing, and a marketing site in seven languages.

  • Built the world's first general-purpose ontology OS alone, from research through to commercial release: roughly 468,000 hand-written lines, seven languages, signed macOS and Windows builds. Solo · shipped commercially
  • Measured all 500 SWE-bench Verified tasks on the official harness, and isolated the scaffold's own contribution at 54% → 76% with the model and a 50-task set held fixed. 78.80%
  • Ranked first on the official DP-Bench leaderboard for table structure recognition, above Upstage, AWS, LlamaParse, Microsoft and Google. TEDS-S 98.11
  • Diagnosed a failure mode reported in the public literature and applied CSLS to retrieval seeding, then repaired a loss that only appeared when the measurement was repeated at 300× the scale, at no additional API cost. recall@10 77.5 → 97.0
  • On the public MuSiQue multi-hop benchmark the graph arm retrieved all supporting paragraphs for 63.0% of questions against 23.0% for a keyword-only arm. Of 17 four-hop questions, none succeeded without the graph. p=2.5e-07
  • Null results and failures to reproduce go on the public record: a fixture-level gain that did not survive a 3,002-note vault, a p-value that moved from 0.038 to 0.478 on a second run of the same design, and a shipped commit reverted after an oracle comparison. p 0.038 → 0.478
  • Wrote the physics engine for the 3D graph view, running half a million nodes under live physics. 500k · 18.3ms/tick
  • Received an acquisition offer from Wrtn Technologies and declined it to keep developing independently.

Rust · TypeScript · React · Tauri · Oxigraph · RDF/SPARQL · WebGPU · Supabase

Projects

Consilience

Seoul

Freelance and personal projects

Full-stack engineer — Anthropic and OpenAI hackathons

Over five months I designed, built and shipped two products alone. The second, Rxly.ai, reached the final round of Anthropic's hackathon.

  • Selected for the final 230 of roughly 13,000 applicants in Anthropic's Built with Opus 4.6 Claude Code hackathon, and the only Korean participant. 230 of 13,000
  • Implemented FFT, formant extraction and voiced/unvoiced detection in the browser, holding phoneme-level error under 10 Hz. ±10Hz
  • Made time-axis alignment 18× faster with a hybrid DTW, with inversion error held at zero. 422ms → 23ms

TypeScript · Next.js · Web Audio API · DSP · LLM Agents · FHIR R4

Projects

Rxly.ai Xhadow

Seoul

Intergalactic Inc.

Founder and CEO · Strategy team

I built LOCL, an all-in-one operations platform for class and studio businesses, and ran the company for three years. As founder I handled the product, design, engineering, marketing, fundraising, legal, tax and hiring myself, and also served as PO for a team of more than ten full-time staff.

  • Raised venture capital, central-government R&D grants, public support funding and guarantee-backed financing. ₩1.4B+ raised
  • Graduated 23rd of 100 companies in cohort 5 of the Changgu programme run by Google and the Ministry of SMEs and Startups. 23rd of 100
  • Designed and built the whole surface — customer app, admin back office and CMS: 1,390 dynamic screens across 1,042,300 lines. 10,000+ users · ₩100M+ GMV
  • Opened in two Shinsegae Starfield locations and Hanwha Galleria, and ran the pop-up store programme.
  • Pitched at demo days hosted by Google, Microsoft, JP Morgan and the UK's YBI.
  • Served as director of the company's in-house research institute under the Startup Growth Technology Development programme, a central-government R&D grant.

TypeScript · Next.js · Nuxt.js · Vue.js · React · Node.js · AWS · GCP · Docker

Projects

LOCL Vime

Seoul

Hansol PNS

Senior Researcher, Standard Software Research, ITS Division

I designed, built and rolled out the in-house standard framework for the company's manufacturing execution system, and left to start my own company.

  • Proposed and built a setup that put responsive web, iOS and Android in one monorepo with Expo and Next.js, in order to cut development and maintenance cost.
  • Wrote and published the standard library and its guidelines, and trained the developers who used it.

Next.js · Expo · React Native · TypeScript

Projects

MES Cross-platform Standard Framework

Seoul

Intro

Software outsourcing and in-house product development

I worked on client projects and the company's own products in parallel.

  • A property transaction management platform (client work)
  • Handset identification and accessory locator software (client work)
  • A personality-quiz marketing platform (in-house)

JavaScript · TypeScript · Node.js

Seoul

Sewon Atos

Researcher, R&D Institute

I built the company's standard ERP framework, then designed and implemented — alone — a vision model that classifies apparel categories whose definitions differ from storefront to storefront.

  • Trained a CNN on 32,474 crawled images with a single GPU to classify apparel categories automatically. 90%+ accuracy
  • Wrote the vector-similarity data cleansing and post-processing pipeline myself.
  • Handled the EWERP research and development that moved a Delphi-based ERP onto a cross-platform web stack.

Python · TensorFlow · Keras · CNN · JavaScript

Projects

Garment Type Classification Vision AI

Skills

  • AI Agent Systems

    Expert

    Agent loops · tool contract design · MCP client and server (OAuth 2.1 · PKCE) · GraphRAG · Personalized PageRank · CSLS · Offline evaluation harnesses · Computer use · approval gates · threat modelling

    Consilience Wrtn Technologies Vime Xhadow Rxly.ai

  • Founding & Operating

    Expert

    Venture fundraising · government R&D grants · guarantee-backed financing · PO for a team of 10+, and the collaboration system it ran on · Legal, tax, hiring, and running a registered corporate research lab · Figma · Framer · brand identity · GA · Meta API · GTM instrumentation and reading the numbers

    Intergalactic Inc. LOCL

  • Ontology & Knowledge Graphs

    Expert

    RDF / W3C quads · named graphs · SPARQL 1.1 · Oxigraph · RocksDB · SKOS · Dublin Core · schema.org · PROV-O · GeoSPARQL · Entity resolution (union-find · embeddings · LLM adjudication) · Louvain / Leiden community detection

    Consilience Wrtn Technologies

  • Product & Full-stack

    Expert

    TypeScript · JavaScript · Python · C/C++ · C# · Java · Next.js · Nuxt.js · Vue.js · React · React Native · Expo · PostgreSQL · Supabase · AWS · GCP · Docker · UX design for ERP, CRM and LMS operational screens

    Consilience LOCL Vime MES Cross-platform Standard Framework Garment Type Classification Vision AI

  • Research Methodology & Benchmarking

    Expert

    Pre-registered hypotheses, rejection thresholds and the commit hash at registration · Paired statistical tests implemented directly (McNemar · Fisher · Wilcoxon · Holm-Bonferroni) · Same-arm controls to establish the noise floor; pooled replicates to test whether significance reproduces · Blinded LLM judge panels, false-negative controls, and deterministic scoring with no judge · Adversarial refutation rounds; null results and failures to reproduce on the public record; do-not-re-propose lists · Public benchmarks measured first-hand: SWE-bench · MuSiQue · DP-Bench · OmniDocBench

    Consilience Garment Type Classification Vision AI

  • Systems & Performance

    Advanced

    Rust (rayon · Barnes-Hut · int8 quantisation) · WebGPU / TSL compute shaders · tree-sitter code graphs · Tauri desktop · code signing · auto-update · Embedded · UART · MIDI

    Consilience INTRO — a piano teaching instrument

Publications & Talks

  • Method

    Diagnosing dense-retrieval hubness with a 2010 distance-concentration result, and porting a cross-lingual correction into retrieval

    Three intuitive alternatives were tried and all three rejected on measurement. The fix came from another field and cost nothing extra in API spend. Re-measured at 300× the scale it turned out to lose in one place, and the repair is recorded here too.

    • 77.5 → 97.0 dense seeding recall@10; keyword-only scored 85.2 on the same set
    • $0 added API cost; end-to-end answer accuracy 87% → 97%

  • Method

    A defect audit with a refutation round — from 45 candidates to 14 reproducible defects

    Suspicions were turned into runnable reproductions, every defect was paired with a control arm, and independent skeptics tried to refute each one; only survivors were recorded as defects. The largest finding was not on the plan, and surfaced only because the audit read a live installation.

    • 45 → 14 candidates to reproducible defects; 14 of 14 survived refutation
    • 102,020 quads lost silently, found by the audit and now covered by a regression test

  • Benchmark

    Two document-parsing benchmarks — first on one, mid-table on the other, and why

    A general-purpose VLM with a prompt on top, not a fine-tuned document parser, took first place on DP-Bench table structure. The same pipeline is mid-table on OmniDocBench; the loss is diagnosed, and a scoring artifact is traced down to the 49 documents that produced it.

    • 98.11 DP-Bench table structure TEDS-S, first on the official leaderboard
    • 79.23 the same pipeline on OmniDocBench Table-TEDS; the leader scores 93.42

  • Method

    Extracting more does not move retrieval — a null result, and a significance that did not replicate

    A paired A/B asked whether a gain observed on fixtures survives in a 3,002-note vault. It did not, and the one effect that did appear lost its significance on an independent repeat. Both results are on the record, and the citation rule changed as a result.

    • +0.1pp retrieval gain bought by 17% more relations (p=0.84, pooled over five runs)
    • 0.038 → 0.478 the p-value on two independent runs of the same design and corpus

  • Method

    Measuring a shipped commit, then reverting it

    One commit that shipped without measurement was measured afterwards. Two of its three changes lost or bought nothing, and all of them were reverted. A separate defect the experiment exposed was left unfixed on purpose: its measured blast radius is nine relations.

    • 0 / 3 times the new implementation matched the oracle; reverted under a rule fixed in advance
    • 0 / 60 calls the tolerant parser actually recovered; the failure it targeted was out of its reach

  • Method

    A default-model swap changed the language documents were written in, and the prompt could not fix it

    A user report that documents came out in the wrong language was traced to a default-model swap and reproduced across six cells. Five prompt wordings each fixed one case and broke another, so the decision the model could not make was moved into code.

    • 20/20 vs 8/20 documents that kept the source language, before and after the swap (Fisher p=4.5e-5)
    • 0 / 5 prompt wordings that passed every cell; moving the decision into code then passed 160 of 160

  • Benchmark

    MuSiQue multi-hop QA — the graph is a precondition, not an improvement

    The graph condition and a keyword-only condition, compared on the same question set. The gap widens with the number of hops, and at four hops the control arm scores zero.

    • 63.0% vs 23.0% questions with every supporting paragraph retrieved (ALL@10), graph vs keyword-only
    • 0.0% of 17 four-hop questions, none succeeded without the graph

  • Benchmark

    SWE-bench Verified across all 500 tasks, and what that number cannot claim

    The result of running all 500 tasks on the official harness. That run cannot separate the scaffold's contribution from the model's, so the separation was measured in two controlled experiments instead.

    • 78.80% 394 of 500, official harness, no retries
    • 54% → 76% model and 50-task set fixed, selection procedure swapped

Education

  • Seoul, South Korea

    Seoul Youth Startup Academy, cohort 14

    Completed · Technology entrepreneurship

    • A national startup programme run by the Ministry of SMEs and Startups and the Korea SMEs and Startups Agency (KOSME)
    • Youngest admitted to the cohort; completed the full one-year programme
  • Seoul, South Korea

    Dongyang Mirae University

    B.Eng. · Computer Software Engineering

    • Exhibited in the IT category at KES (Korea Electronics Show) as team lead, 2019 and 2021
    • Taught at the university as a startup mentor after graduating; received a plaque of appreciation in 2024

Languages

  • Korean

    Native

  • English

    Fluent — professional working proficiency

Honours & Activities

Awards

  • Dongyang Mirae University — plaque of appreciation

    Dongyang Mirae University

    Awarded for serving as a startup mentor.

  • Incheon CCEI investment demo day — excellence award

    Incheon Center for Creative Economy and Innovation

  • Startup Experts — first place

    Y&Archer and six partner organisations

  • COEX University EXPO — grand prize

    COEX

  • (date not recorded)

    Seoul Campus Town resident-company contest — award

    Seoul Campus Town

  • (date not recorded)

    Technology startup idea competition — award

    (to be filled in — organiser)

  • (date not recorded)

    TensorFlow Korea — Rising Star

    TensorFlow Korea

Funding & Guarantees

  • More than ₩1.4B raised in total

    Venture capital, central-government R&D, public grants and guarantees

  • Two venture rounds closed, at a valuation above ₩1.5B

    One Billion Partners · Real Vision — two firms across three funds

  • (date not recorded)

    ₩500M in R&D and youth-startup guarantees

    Korea Technology Finance Corporation · KOSME

Programmes & Selections

  • Anthropic 'Built with Opus 4.6' Claude Code hackathon — finalist, the only Korean

    Anthropic · Cerebral Valley

    About 13,000 people applied, 500 were selected to take part, and I was one of the 230 who reached the final round. The entry was Rxly.ai.

  • High Flyer investment track — accepted

    JP Morgan · Work Together Foundation

  • Seoul Youth Startup Academy, cohort 14 — youngest admitted, graduated

    Ministry of SMEs and Startups · KOSME

  • Startup Success Package fundraising accelerator — final selection

    Smilegate Investment

  • Google Changgu programme, cohort 5 — graduated 23rd of 100 companies

    Google · Ministry of SMEs and Startups · KISED

    The programme supports Korean app and game startups.

  • Early-stage startup support programme for the arts — final selection

    (to be filled in — organiser)

  • Startup Growth Technology Development grant — also served as R&D lab director

    Ministry of SMEs and Startups

    This is a central-government R&D programme.

  • Preliminary Startup Package — youngest admitted, graduated with distinction

    Ministry of SMEs and Startups · Incheon CCEI

  • (date not recorded)

    Selected for and pitched at demo days run by Google, the UK's YBI, Microsoft and the KEPCO Foundation; cited as an exemplary presentation

    Google · YBI (UK) · Microsoft · KEPCO Foundation

  • (date not recorded)

    Hanyang University Innovation Startup School — selected

    Hanyang University

  • (date not recorded)

    Innost Ground Challenge — accepted

    Chungnam Information & Culture Industry Promotion Agency · Zephyrus Lab

  • (date not recorded)

    Seodaemun-gu local venture programme — selected

    Seodaemun-gu, Seoul

Certifications & Ratings

  • (date not recorded)

    Rated an excellent technology company by NICE

    NICE Information Service

  • (date not recorded)

    Certified as a venture business

    Ministry of SMEs and Startups

Business results

  • (date not recorded)

    Supply agreements with Shinsegae Starfield and Hanwha Galleria; ran the in-store pop-ups

    Shinsegae Starfield · Hanwha Galleria

Scholarships

  • (date not recorded)

    Scholarships from four foundations and institutions, awarded on direct application

    Hyundai Motor Chung Mong-Koo Foundation · Seoul Future Talent Foundation (formerly Seoul Scholarship Foundation) · Korea Exchange (KRX) · Seoyon Group's Cheonjeong Foundation

Speaking, Judging & Mentoring

  • Korea Youth Startup Competition — judge and guest speaker

    Ministry of Education

  • (date not recorded)

    Selected for the 1,000-strong career mentor corps

    Incheon Cyber Career Education Center

  • (date not recorded)

    Guest lecturer and startup mentor

    Incheon Metropolitan Office of Education · Seowon University · Dongyang Mirae University

Exhibitions

  • Korea Electronics Show — IT category exhibit, team lead

    Korea Electronics Show, COEX

  • Korea Electronics Show — IT category exhibit, team lead

    Korea Electronics Show, COEX

  • Future of Education Conference — IT category exhibit, INTRO

    Ministry of Education, KINTEX

  • Korea Electronics Show — IT category exhibit, INTRO, team lead

    Korea Electronics Show, COEX

Patents & IP

  • Two patents and one trademark filed

    Korean Intellectual Property Office

    (to be filled in — titles and application numbers)