What the customer asked for
“Our employees cannot use SQL. Help them access and use the data.”
Major Korean logistics enterprise, 200-user beta two weeks after I joined.
DongHyuk LimForward Deployed Engineer
I find the problem behind the request and build the system that solves it.
I work with customers to understand their workflows, define what is needed, and own the path from design to deployment.
How a request becomes a system
Not every request survives contact with the workflow. The ones that do come out specified.
Request → problem → system — Three projects, one pattern. Pick a case.
What the problem turned out to be
Query syntax was not the first barrier. Operators did not know what data existed, where it lived, or what to search for.
Found by working on-site and by using competing data-search products to locate the real cause.
What I built
A discovery layer over governed Databricks data: hierarchy, filters, result context, a query builder and an operational governance workflow.
Beta completed with 200 users. Governance workflow still in development.
200-user betaOn-site FDEDatabricks
Read the case study
What the customer asked for
“Build an MCP server for investors.”
No specification behind the request — a technology name, not a product.
What the problem turned out to be
The useful scope sat where three things overlapped: what investors actually did, what the platform data could support, and the security model already in place.
3 VC investors interviewed, 115 platform endpoints checked against real tasks.
What I built
An MCP server exposing six investment workflows through the customer’s own login, with explicit per-company context and data isolation.
Two months, solo, delivered as a paid customer project.
2-month solo delivery6 workflows115 APIs analysed
Read the case study
What the customer asked for
“Generate two related questions from the same passage.”
A hard requirement for embedding our tool inside the customer’s own product.
What the problem turned out to be
The pipeline generated one independent item per pass and had no passage-level context. Another prompt call would produce two questions, not a valid linked set.
I separated what could improve immediately from what needed a new architecture, and said so to the customer.
What I built
A passage-centred pipeline: staged generation, level-aware retrieval, multi-stage validation, reworked token management and cache.
Shipped in three weeks. 15s → 8s, 5% → 1.5% failure rate.
3-week redesign15s → 8s5% → 1.5%
Read the case study
0users
Enterprise beta completed with me as the on-site FDE
Case 01
0months
From an undefined MCP request to a delivered product, solo
Case 02
0weeks
Unsupported requirement redesigned, implemented and shipped
Case 03
15s8s
Generation latency after the pipeline redesign
Case 03
01
Career
Four roles.
Four roles.
The same distance from the customer.
Founder, engineer, operator, forward deployed — the seat changed, the job did not: define what is actually needed, then build it.
-
Mar 2026 — Present Current
Upflow
Forward Deployed Engineer
Embedded with enterprise customers as tier-2 delivery under a prime SI. I take the request the customer brings, find the problem behind it, and own the system through to the workflow the end user touches.
- Enterprise Data Portal Data-discovery layer over Databricks for a major Korean logistics enterprise
- Investor-Relations AI Assistant Enterprise agentic-AI engagement: AX consulting and MCP architecture
~350M records surfaced200-user pilot shipped6 MCP workflows scopedUSD 7K AX contract -
Mar 2025 — Jan 2026
InJob
Founder · Product Owner
AI career-recommendation startup for university students. I ran the discovery myself — 80+ student interviews — and turned it into a curriculum-based recommendation product that addressed career-information search inefficiency.
USD 17K Korean Gov. Pre-Startup Package200-student validated MVP86% PoC satisfactionOCR pipeline 5 min → 30 s -
Feb 2024 — Dec 2024
HyperLearning
Engineering & Operations
B2B English question-generation service on LangChain and RAG. I redesigned the pipeline around a customer requirement the existing architecture could not support, and rebuilt retrieval, validation, caching and fallback handling around it.
USD 40K annual B2B contractLatency 15s → 8sFailure rate 5% → 1.5% -
Aug 2023 — Dec 2023
Youngsung Kim’s AI Research Group
Undergraduate Research Intern
Researched transformer-based architectures and reviewed state-of-the-art literature in the field.
02
Selected work
Three customers.
Three customers.
The same job in three different systems.
Each case study documents the request, the problem behind it, the constraints, the decisions, the architecture, what shipped, and what is still unfinished.
01 — Logistics · Enterprise data
Enterprise Data Discovery
Helping operations teams find and use governed data without SQL
Reframed a request for “easier SQL” as a data-discovery problem, joined on-site two weeks before beta, and completed a 200-user enterprise beta.
Read the case study02 — Investment platform · Integration
Investment Workflow MCP
Turning an MCP request into six working investment workflows
Interviewed 3 VC investors, mapped 115 APIs to real tasks, and delivered the MCP server, customer login and multi-company isolation solo in two months.
Read the case study03 — Education · AI systems
AI Generation Pipeline
Redesigning a production pipeline around a requirement it could not support
Explained an architectural limit to the customer, scoped it into a three-week redesign, and rebuilt generation around shared passage context.
Read the case study
03
How I work
Not a list of skills.
Not a list of skills.
Four things that repeated in every project.
Each one is followed by the moment it actually happened.
01
Start with the workflow
I study how people actually work before deciding what to build.
On-site with logistics operators, I watched people stall before the query screen, not on it. The product became a discovery layer instead of a SQL helper. Case 01 — Problem discovery →
02
Find the problem behind the request
The initial request is evidence, not automatically the specification.
“Build an MCP server” named a technology. Three investor interviews and 115 endpoints turned it into six workflows worth building — and removed the ones the data could not support. Case 02 — Discovery →
03
Design for the real environment
Existing systems, data boundaries, ownership and operational constraints shape the solution.
One investor can belong to several companies, so company context — not just authentication — had to travel with every tool call. The customer’s own login stayed authoritative. Case 02 — Authentication and company switching →
04
Build, test, and revise
I implement the system, test it with users, and rebuild when the first answer does not work.
A results screen that returned correct data still failed, because users could not read it. I rebuilt the information architecture — and separately, ran ~800 GPT API tests to make pipeline decisions on repeated behaviour rather than single examples. Case 01 — Implementation · Case 03 — Testing →
04
Why FDE
A new domain every time.
A new domain every time.
The same job underneath.
Data-discovery tools for logistics operators. Investment workflows wired into an enterprise’s own authentication. AI generation pipelines for education. Every domain arrived with different users, different constraints, and a different definition of “good”.
What repeated was the work itself: learn a new domain quickly, work out the service the customer actually needs, and build it so it runs inside their environment. That is why I want to work as a Forward Deployed Engineer.
Also delivered
Korean writing assessment1,500 concurrent candidates · 4,500 answers
With 1,500 people sitting the exam at once, an answer that failed to save invalidated the sitting — the tolerance for error was effectively zero. I built the backend around traffic and error handling and added live log monitoring over the runs.
The rubric existed only as a professor’s handwritten marks. A random forest over nine engineered features — sentence length, lexical variety, word difficulty among them — scored above 80% on the held-out set against 70% for human raters. Its training data came from 6,000 handwritten scores digitised through an OCR pipeline.
Presentation analysis AIHour-long recording: 30 min → 3 min
Scoring something as subjective as delivery meant inventing the rubric first. Intonation became pitch variation; pronunciation became the diff between the recognised text and the original script.
Those measures feed an automatically generated speech-behaviour report. The build stopped at demo quality and never went in front of real users.
Awards & selections
Grand Award — Project Marathon Hackathon1st of 53 teams · 2023
Sole owner of the speech AI track in a 7-person team. Off-the-shelf Whisper missed syllable boundaries in presentation audio, so I cleaned Korean speech data against inter-word syllable recognition and fine-tuned a model for the setting.
Analysing an hour-long recording took 30 minutes. Splitting on silence — the points where the signal falls to zero — kept words intact where arbitrary cuts would not, and processing those segments in parallel brought it down to 3 minutes.
Selected — Startup-Centred University ProgrammeUSD 43K in funding · 2026
Revenue of roughly USD 14K and no outside investment made the quantitative case weak, so the application had to win on reasoning rather than numbers.
I crawled this year’s government programme announcements to work backwards to what reviewers were funding, then found that all three AX projects the company had delivered were logistics firms — which became the story: helping mid-sized logistics companies that fail at AX because they cannot hire developers. Sole author of the ten-page plan.
Excellence Award — InnoThink Makerthon3rd of 9 teams · 2023
Operators were hanging helmets on the scooters and losing them. Rather than treating helmet loss and illegal parking as two problems, I put a folding helmet locker in a docking bay where the fare only settles once the scooter is docked.
Determined the mounting position from the scooter’s geometry first and fitted the software to the hardware, reaching a working Raspberry Pi NFC prototype inside 24 hours.
Placed — Student Ship Design ContestSNAK · 15 universities · 2023
Led a 20-person team designing an AFRAMAX crude oil tanker compliant with EEDI Phase 3. No standard for optimal dimensions existed, so I defined resistance minimisation as the criterion, built a dataset of vessels in service, narrowed the principal dimensions by regression, then implemented the Holtrop-Mennen method in Python to estimate the minimum-resistance form.
When the fluids track slipped near the deadline and blocked structures entirely, I merged the two tracks to run in parallel rather than holding to the sequence, and we finished on time.
Excellence Award — Ocean Engineering CAE ContestKSOE · 2022
Designed a floating breakwater with a new cross-section for long-period wave attenuation. Holding volume and ratio equal to the box-type baseline kept the comparison fair, and the new form attenuated 33% more wave energy.
Ran Hypermesh structural analysis and OrcaFlex mooring assessment across normal, typhoon and extreme sea states; peak effective tension of 2,600kN stayed within the 4,940kN allowable.
- Leadership ScholarUniversity of Inha · 2023
05
Contact
I am interested in engineering roles where understanding the customer is part of building the product.
Static site, no contact form. Email is the fastest route.