Ilkyu Lee

AlphaRPA Inc. — CEO

2022.07 – Present · Founder & Product Owner

Founded, launched, and operated alphanote.io, a trading-journal analytics platform. Owned the whole product — customer problem definition, data pipelines, AI analytics, and user-facing reports.

• Raised over ₩600M in government grants and funding
• Top-rated selection in the Pre-Startup Package; Data Voucher program; KODIT (Korea Credit Guarantee Fund) guarantee
• Two Didimdol R&D grants — LangChain-based trading-skill analysis AI (lead) / BERT-based expert–company matching (subcontract)
• Hanam Venture Center (2022) → moved into TIPS Town (2023 ~)

• Problem — individual investors struggle to analyze their own trading data systematically
• Built the data pipeline — tick data for 2,300 tickers, 20GB/day, Kafka → InfluxDB, 5TB accumulated
• Automated collection of trader records — 15M records accumulated
• Built an AI research environment (feature selection, performance testing, visualization, task manager)
• Reassessed direction as market conditions shifted; decided to pivot the product

• Problem — traders repeat mistakes because it is hard to see why a trade won or lost
• Designed ~70 FastAPI endpoints; OAuth and payment integration
• RAG-based trade coaching — feature normalization → Sentence Transformer embeddings → vector store → similar-trader retrieval → LLM analyzes differences in sizing and timing to generate advice
• Acted on user feedback — cut displayed information to 20%, layered the UI, added automatic journal upload
• Reached 2,000 DAU through community channels
• Retention fell short of target → promotion experiments (no significant lift) → decided to scale down after profitability analysis

BrainSoft — AI Researcher

2020.03 – 2021.08 · Speech AI

• Built GPU servers and IDC infrastructure starting from no ML environment
• Set up Korean speech-recognition training on ESPnet — 1,000 hours of training data
• Validated the company's patented spectrogram features — millisecond-level, 100× the previous scale
• Implemented an attention-based speech-to-phoneme model (ARSG) directly from the paper
• Early-stage architecture and technology decisions; ESPnet open-source contributions

AI Research Institute (AIRI) — AI Researcher

2018.02 – 2020.02 · Computer Vision · Finance Data Automation

• Multimodal avatar — integrated four separately-run vision modules into a Redis-based system; three field demos (ETRI medical consultation, public signage, smart-home exhibition)
• Person re-identification — within 3% of published accuracy; ported from PC to Jetson at 20FPS. Hyundai Motor collaboration
• Finance systems automation · data cleaning — web-data RPA collection (UiPath, JS injection), data cleaning and pattern analysis, ELK logging system for the finance AI team
• Designed the company's internal coding-test program and grading scheme
• (Intern) MTCNN-based face detection/alignment, age regression and gender classification. Hyundai Motor collaboration

LLM Auto-Tagging · AI Content Generation Solution

SME Technology R&D (Didimdol) · ₩160M project · Technical consulting & development

Integrated inventory-management solution for a fashion e-commerce company — built the three core modules (auto-tagging, content generation, real-time inventory sync) and wrote 68 electronic lab-notebook entries.

• Auto-tagging — conditional ontology by gender and category, two-stage prompt chain, logic guardrails, LoRA fine-tuning
• Content generation — image captioning, per-channel copy generation, an output parser that cross-checks tags to filter hallucinations
• Real-time inventory sync — at-least-once message queue with DLQ, write-invalidate cache, sync engine
• Baseline measurements (before fine-tuning) — tagging Macro-F1 78.5%, inventory-sync accuracy 95%, API 23ms, 44 img/s. Low color/season accuracy → identified the need for multimodal training

Store Operations Intelligence SaaS

TIPS project · ₩800M project · Product design & TIPS pitch

For boutiques and pop-up stores — combines de-identified CCTV movement data, POS, and Naver market data to tell owners what to do today to lift sales. Owned everything from product definition through the R&D proposal and the TIPS pitch.

• Documents before code — fixed 13 spec documents up front: domain, personas, DB schema, 16 insight definitions, action catalog, LLM prompts, phase 1–4 roadmap
• Action catalog — every recommendation ships with expected impact and evidence; "shop-owner language" principles and a glossary translate the jargon
• R&D proposal — built on group inference + an operations knowledge graph, restructured around implementation for reviewers
• Planned stack — FastAPI, React + TypeScript, PostgreSQL (pgvector), Redis, LLM API, Naver Search API, Docker Compose

Payroll System — Korean labor-law firm

2021.11 – Present · 5 years

A five-year engagement with a single client — carried the system through a full generational change from desktop to web.

• Gen 1 — PyQt5 desktop suite. Company/employee/payroll management; onboarding, overtime-pay, and revenue aggregation. MongoDB Atlas, shipped as executables running in their office
• Gen 2 — rewritten as a React + FastAPI + MongoDB web service. Upload a payroll ledger with 3-row headers → column names assembled and types inferred to build the DB; per-company 3-year status at a glance

Work Automation — consumer-lending company

2025.08 – Present

Automated systems gated by authentication and regulation, deployed into daily operations.

• Collection-account inquiry automation — brokerage login (encrypted credentials), account and transaction collection, incremental queries computed from the last-fetched date, staff GUI. Verified in production
• Court e-filing service automation — certificate login → case list from Excel → correction/dunning lookups → staged first re-service and second special service → results logged. Module-level tests
• Bulk call-recording collection tool

AI Products

2025 – 2026

• AI report generator — Stanford STORM methodology. PDF upload → multi-perspective question generation → research → outline → sections → automatic citations and bibliography. FastAPI + Firebase + GPT-4 / React + TypeScript
• Clinic-inventory Decision OS — ERP integration, Next.js + TypeScript. Anomaly, dead-stock, expiry, order-planner, reorder, vendor, and trend dashboards plus natural-language inventory queries. Built the MVP skeleton and handed it to the team
• KakaoTalk-channel AI-agent integration — Korean localization of a global Web3 company's AI agent. Solved Kakao's 5-second response limit with a callback API structure. Worked with HQ directly in English; 10 technical questions reflected in the final proposal
• Court-auction appraisal-report collection tool for a real-estate firm