DongHyuk Lim

Enterprise Data Portal — data-discovery layer over Databricks for a major Korean logistics enterprise; tier-2 delivery under the prime SI

• Reframed the request from a SQL-training problem into a data-discoverability problem, then owned planning and development of the discovery layer over a partner-maintained Databricks platform — surfacing ~350M records to non-technical operators.
• Rebuilt the search-results experience end to end after early users said it was not communicating — benchmarked data-search SaaS to isolate hierarchy and filtering as the drivers of comprehension.
• Designed the data-governance feature set and ran delivery as PM across the prime SI and the end customer; prototyped a natural-language Query Builder PoC on the Databricks LLM API; shipped to a 200-user pilot.

Read case study 01 →

AI career-recommendation startup for university students

• Designed an OCR-based metadata pipeline from user feedback on manual curriculum input, reducing the process from 5 minutes to 30 seconds.
• Translated insights from 80+ student interviews into a curriculum-based career-recommendation algorithm, addressing career-information search inefficiency and achieving 86% PoC satisfaction.
• Selected for the USD 17K Korean Government Pre-Startup Package; validated the MVP with 200 students.

B2B English Question Generation AI Service — LangChain/RAG content-generation service for an enterprise customer

• Redesigned a LangChain/RAG pipeline for an enterprise customer’s “1 passage, 2 questions” requirement, helping secure a USD 40K annual B2B contract.
• Improved GPT generation via prompt engineering, retrieval, caching, validation and fallback handling — latency 15s → 8s, failure rate 5% → 1.5%.
• Built the backend and scoring model for Konkuk University’s Korean writing assessment — 4,500 answers from 1,500 concurrent candidates across two semesters, delivered with no incident in either run. A random-forest model over nine engineered features scored above 80% on the held-out set against 70% for human raters; digitised 6,000 handwritten scores via an OCR pipeline to train it.

Read case study 03 →

• Researched transformer-based architectures and reviewed state-of-the-art literature in the field.

• Relevant coursework: Numerical Analysis, Machine Learning for Mechanics, Data Structures.

Scoring how someone presents, not just what they said

• Off-the-shelf Whisper is trained on general Korean and missed syllable boundaries in presentation audio, so I collected Korean speech data, cleaned it against inter-word syllable recognition, and fine-tuned a model for the presentation setting.
• Analysing an hour-long recording took 30 minutes. Splitting on silence — the points where the signal falls to zero — kept words and sentences intact where arbitrary cuts would not, and processing those segments in parallel brought the run down to 3 minutes.
• Turned subjective delivery into numbers a model could output: intonation as pitch variation, pronunciation as the diff between the fine-tuned transcript and the original script. The measures feed an automatically generated speech-behaviour report.

A new cross-section for long-period wave attenuation, verified through structural and mooring analysis

• Selected the installation site by analysing Korea Hydrographic and Oceanographic Agency sea-area data, then derived the principal dimensions from published formulae while holding volume and ratio equal to the box-type baseline so the comparison stayed fair.
• The resulting cross-section attenuated 33% more wave energy than the box-type baseline in simulation.
• Ran the Hypermesh structural analysis and the OrcaFlex mooring assessment across normal, typhoon and extreme sea states; peak effective tension reached 2,600kN against a 4,940kN allowable. RAO analysis was a teammate’s work, which I reviewed.

Helmet loss and illegal parking treated as one problem rather than two

• Operators were hanging helmets on the scooters themselves, which kept losing them. Behavioural research on riding patterns pointed at a single structure instead: settle the fare only when the scooter is docked in a designated bay, and put the folding helmet locker in that same bay.
• Determined the mounting position from the scooter’s own geometry first, then modelled the parts and fitted the software to the hardware rather than the other way round — a working 3D-printed prototype on Raspberry Pi inside the 24-hour window.

Validating demand offline before spending anything on production

• Ran a one-day booth on a researched dog-walking route and recorded roughly 40 on-site responses covering the dogs’ reaction, purchase intent and acceptable price. Willingness to pay came in above our estimate, so the product moved upmarket to a premium line.
• Turned 15 leads from that booth into an Instagram channel that reached 1,000 followers in about two weeks, with DMs routing interest back to offline tastings. The crowdfunding round closed at 107% of goal.

• Organised summits and workshops, fostering a community of over 200 students.
• Mentored students on product management and competitive programming techniques.
• Designed and ran the chapter hackathon end to end — a month-long run of ideation, overnight build, demo day and finals for 50 participants across 10 teams, with a 14-person organising team. I added the demo day session so teams would meet real users before finishing; a token-based investment mechanic pushed feedback quality, and 95% of visitors rated the session positively. Other chapters have since reused the format.
• Wrote and pitched the sponsorship proposals that brought in three sponsors — the SW-Centred University Programme, LINC 3.0 and Google — worth USD 1.4K in kind plus awards issued in the programmes’ names.

• Delivered AI-assisted PoC development tech talks to 50+ participants per session, covering prototype-to-deployment workflows.
• Mentored hackathon teams on technology selection and solution architecture for deployable AI PoCs.