Janghoon Lee
Technical co-founder and CTO at Redrob since 2018. Built the engineering organization from zero across three countries, and still write the code I ship.
Experience
Chief Technology Officer · Redrob
Mar 2018 - Present
Teams in Seoul, India, and the US
$6.8M ARR (Nov 2025) · $14M raised
- Shipped Redrob as a low-cost LLM and careers platform for emerging markets: Chat, Jobs, Skills, and Signal under one brand, free campus access under a fixed per-student cost ceiling, and $6.8M ARR (Nov 2025) with 10M+ user signups across 500 universities in India.
- Built the engineering organization from 0 to 50+ across Seoul, India, and the US: hiring bar, review process, and regional ownership (India on-call for Indic text and Jobs; Seoul for model and infra; US for enterprise Console), while still writing the production code I ship.
- Ministry of Science and ICT minister commendation at the 2024 ICT Fund awards and results conference (cybersecurity) for developing and operating Redrob, a global recruiting AI platform built on AI–blockchain jobseeker identity and capability verification.
- Owned model and cost constraints end to end: routed Indic text (Console), GEPA eval harness (Apache-2.0), self-hosted Image and Tune paths, and Studio’s dual LLM/VLM safety gate (99.92% accuracy, 100% recall on 6,000 prompts).
Education
Minerva University
M.S., Decision Analysis · Sep 2022 - Aug 2024
Credit Bank System, National Institute for Lifelong Education (NILE)
B.Eng., Computer Engineering · Feb 2022
Minerva University
Computer Science · 2018 - 2021 (founding leave until 2021)
Projects
Redrob product and research.
Redrob$6.8M ARR (Nov 2025) · 10M+ signups · 500 universitiesTechnical co-founder and CTO since March 2018. The platform reached $6.8M ARR and 3 million users across 500 universities in India as of the Nov 2025 Series A, and later 10M+ user signups, built by an engineering organization spread across Seoul, India, and the US.
Redrob EvalPublic harness, Apache-2.0LLM eval workbench: Compare, Evolve, Deploy. Blind brackets become routing labels. Test spent once. Relative cost only. Apache-2.0, clone and run.
Redrob ConsoleRouted Indic text APIClient-facing API dashboard for a routed Indic text stack: Sarvam’s open model and other open backends behind one Redrob completion surface. I owned onboarding, integration support, and per-customer cost tuning.
Redrob ImageBlind A/B vs GPT Image 2: large gap ruled out, parity not shownOpen-weight Z-Image Turbo (Apache-2.0, commercial OK). ~8 steps, plain Diffusers. Blind vs GPT Image 2: 44-16-30 (n=90). ~1.8s/image. Photo/portrait, weak on text.
Redrob Tune60s track in 14-23s, self-hostedMusic on ACE-Step 1.5 through ComfyUI: Tune Fast and Tune Pro. A 60-second track landed in about 14s on Fast and about 23s on Pro in the Jul 28 preference bake-off. Tune has no standalone release; it ships as the Music tool inside Redrob Studio. Samples below include winners from that run and earlier bake-offs.
Redrob StudioSafety gate: 100% recall on a 6,000-prompt evalLocal Next.js studio for image and audio tools on your own ComfyUI GPU, with dual LLM/VLM safety gates. On a 6,000-prompt NSFW/SFW eval across image, text, and audio: 99.92% accuracy, 100% recall, five false positives. No chat surface and no video generation stack; home is a tools catalog and bake-offs sit in the same shell. Latency and preference claims live on Image and Tune.
Writing
5 Aug 2026
- inference
- on-device
- evaluation
- grammar
Failure Notes: How to Run an LLM on a User Laptop
Users refuse to send documents out, so we ran the model on their own machines, and one document evaluation took 23 minutes on an office laptop. With no GPU, growing the model and shrinking it are both blocked, so we went with a small model, removed JSON, and enforced format with a grammar. Implementing that, we got stuck six times and the cause was our code every time. What raised processing speed by 10x was one option that fetched a full probability distribution every token and then threw it away.
30 Jul 2026
- strategy
- inference
- evaluation
- on-device
Finding the Problems Where Capital Doesn't Decide
My founder friend said our Indic LLM story sounded two years old next to physical AI. The useful work was not chasing a fashionable comparable. It was killing wrong prescriptions until the only remaining space was hardware-constrained inference, evaluation that spends test once, and products that compound from owned usage.
1 Jan 2026
- advice
- entrepreneurship
- startup
Don't Start a Business
New Year brought a wave of questions about whether to found something. I am only nine years in. I do not have much to offer except this: do not. The numbers say stop. The contradiction is that I would start again for the co-founders I had.
Awards
- 2019Google - Best Use of Cloud (2nd place)Infrastructure-as-code deployment of a serverless function that connected optical character recognition (OCR) models to text-to-speech (TTS)
- 2019CrowdStrike - Best Use of Go (1st place)Secure cloud service in Go that converted images to speech in near real time
- 2019USC Business Hackathon - Runner-up (2nd place)Smart glasses prototype that read situational and text images aloud for visually impaired users
- 2018Canon U.S.A. - Semi-finalist (4th place)Proof of concept for open-ended camera storage using IPFS (distributed file network) and the Canon Storage API
- 2018Xingshi Cup Industrial Design/Engineering (3rd place)In-vehicle system using a convolutional neural network (CNN) to detect aggressive driving and suggest slowing down
- 2016Samsung Open Source SW Developer Competition - Silver Prize (2nd place)Youngest recipient in Korea; build pipeline that converted C# into C++ so the result could run on Tizen OS