Inho Lee

Product and seed investing. Currently an MBA candidate at Berkeley Haas.

Looking for a summer 2027 internship in corporate development or product.

Seven years shipping consumer product at Kakao and writing $7.3M of seed checks at D.CAMP. The work I am proudest of came from going and looking — interviewing the agents on a support queue, building the tool rather than asking for it.

Product

Things I specified and mostly built. Open a row; click a screen.

Project Doha

An AI pitch-deck evaluation tool, built at D.CAMP

  • Internal tool
Screening was the fund's bottleneck. An LLM tool that reads every deck against the same rubric and shows its working.
  • Why Roughly 500 decks an intake, about a hundred hours of reading, and two weeks before a founder walks.
  • What An LLM pitch-deck evaluation tool, and a crawler working backwards from rising themes to the companies building them.
  • How Proposed it myself rather than being assigned it; designed the architecture and wireframes, did the API integration hands-on.
  • Impact An estimated 30% off assessment time, the firm's AI Innovation prize against nine entries, and 10+ peer funds benchmarking it.
Input, process and output of the evaluation tool Two inputs: the pitch deck, pulled from the batch database, and company data pulled from an external service by API, covering past applications, financial statements and funding history. The process has three steps: read the deck through OCR, organise what it says against the evaluation sheet for that company's industry, then score every item on that sheet across three criteria, market, competitive position and team. The output is a score for each defined criterion. The sheet was rebuilt with the investment reviewers over repeated rounds of interviews, which raised agreement between the tool's scores and the reviewers' own from about sixty per cent to about eighty. INPUT Pitch deck pulled from the batch database Company data external API: past applications, financial statements, funding PROCESS 1 Read the deck through OCR 2 Organise what it says against the evaluation sheet for that industry 3 Score every item on that sheet Market Competitive Team OUTPUT A score for each defined criterion weighted market 30% · competitive 40% · team 30%, each claim tied to the deck page it came from Agreement with the reviewers' own scores: about 50% → 77% rebuilt with them over repeated rounds of interviews
The sheet is the product, not the model. Rewritten round after round with the reviewers until agreement with their scores went from about 50% to 77%; the model barely changed.
‹Project detail CommitteeHistory In progress ∨
Batch Batch Program, Q4 2026cohort 8
1Project setup 2Company evaluation 3Analysis
Companies500 matched against the company database100%
Document review480/500 56 pass · 24 hold · 400 reject96%
Reviewer meeting0/56 not started0%
IR meeting0/0 not started0%
Final selection0/0 not started0%
Applicant companies
CompanyIndustry ReviewAIStage
Northwind RoboticsMaterials › Robotics5.68.8reject
Calder BioBio › Biotech4.36.8reject
Vellum LabsIT services › Fintech5.35.9hold
Harbor GridClimate › Energy7.98.0pass
Roughly 500 applications an intake, against about a hundred hours of reading. Every company carries two scores — the reviewer's and the model's — so a disagreement is visible in the list.
‹Vellum Labs IT services › Fintech SummaryAI analysisMetrics
AI elevator pitch
MarketReconciliation tooling for small finance teams, growing with the shift to real-time settlement
ProblemMonth-end close is still spreadsheet work at companies below the enterprise tier
SolutionRules engine that reconciles ledger against bank feed and flags only the breaks
ModelPer-seat subscription plus volume pricing above a transaction threshold
AI assessmentrecommend Founders have shipped reconciliation at a larger platform before, and the wedge is narrow enough to be defensible while the category is still forming. 8.2Market · 30% 8.8Competitive · 40% 9.2Team · 30% 8.7Overall
Why now
  • Real-time settlement mandates land in two of the three named markets next year
  • Commercial start in 2025, standards fixed by 2026
Page 2
Competitive position
  • Only package in the deck combining ledger, bank feed and audit trail
  • Two incumbents named, neither below the enterprise tier
Page 3
Market sizing
  • Bottom-up figure given, but capture rate is asserted rather than derived
Page 7
Customer pain
  • Three pilot quotes, no usage figure anywhere in the deck gap
Page 11
Market– Competitive– Team– Total – / 30 PassHoldReject
Every claim carries the deck page it came from. A gap is reported as a gap, not scored as zero — and the reviewer still types their own three numbers.

Claim 2.0

A self-service support architecture, built at Kakao

  • Internal, shipped 2021
Inquiries grew faster than the team and the answer was to hire. Routing by fault let policy absorb a quarter.
  • Why A pandemic surge took inquiries with it; the call-centre partner asked for more agents, which binds cost to demand.
  • What A self-service support architecture that routes each case by whether the customer, the seller or the platform is at fault.
  • How Interviewed our agents and the outsourced ones, built the taxonomy and branching logic, then taught settlement to cover part of a cart.
  • Impact 2,500+ of roughly 10,000 monthly inquiries auto-resolved, automation 0 → 24%, handling time 5 → 4 minutes, leading four people.
Before
Free-text inquiry“one of the three things is broken”
↓
Queued for an agent
↓
No path for part of a cartthe claim is all of it or none of it
↓
Agent unpicks the order by hand
After
Select the itemand the reason, from the order itself
↓
Type classified · fault attributed
↓
Policy resolves itno agent involved
↓
Settlement adjusts for that item only
One item in a three-item order arrives damaged. Before this there was no software path for it, so it queued for a person who unpicked the cart by hand.
Inquiry routed by fault attribution An inquiry is classified into a type, then routed by whether the fault lies with the customer, the seller or the platform. Where claim policy fully determines the outcome the case resolves without an agent; the remainder goes to a person. inquiry classify type taxonomy fault? customer seller platform policy resolves no agent agent
Fault attribution, not the wording of the inquiry, decides whether policy can answer on its own. Simplified here — the real branch tree is wider.
What policy was allowed to decide, and what that moved
CaseBeforeAfter
Damaged on arrival, seller at faultagentpolicy
Wrong item sentagentpolicy
Changed mind, inside the windowagentpolicy
One item of severalno path at allpolicy
Fault genuinely unclearagentagent, unchanged
Auto-resolved per month02,500+ of ~10,000
Automation rate0%24%
Handling time5 min4 min
The fourth row is the one that mattered. Engineering proposed cutting that case, roughly halving the automation rate — I won it back by sizing the volume each feature would absorb.

Seller payout redesign

Moving the settlement cycle from monthly to weekly, at Kakao

  • Internal, still live
Competitors paid sellers weekly; we paid monthly. Paying faster meant teaching settlement to carry a correction forward, not claw money back.
  • Why Competing marketplaces had already moved to weekly settlement, and we were losing sellers to that gap.
  • What A switch to weekly seller payouts, with settlement adjustment that still handles claims and refunds once money has gone out.
  • How Ran cash-flow simulations to size the working-capital impact, then partnered with engineering on the adjustment logic that makes it safe.
  • Impact Completed the redesign before I left Kakao; it has since launched and is live.
Monthly versus weekly payout timing On a monthly cycle a seller waits up to a full month after a sale before being paid, and claims are almost always settled before the payout. On a weekly cycle the wait is days, but a claim can now arrive after the money has gone out, which is what the adjustment logic exists for. MONTHLY sale claim window payout a seller waits most of a month; claims are settled long before the money moves WEEKLY sale payout claim arrives after this is the case the old system had no answer for
Shortening the cycle moves the claim to the wrong side of the payout. Paying faster is easy; paying faster without breaking settlement is not.
Monthly
Sale
↓
Claims and refunds resolve
↓
Net amount paid out once
↓
Seller waitsand leaves for a competitor
Weekly
Sale
↓
Paid out within the week
↓
Claim arrives afterwards
↓
Adjustment against the next cyclerather than a clawback
The fix is to net the correction forward. A claim landing after payout is carried into next week's settlement instead of pulled back out of a seller's account.
What the cycle change moved
MonthlyWeekly
Payout frequencyonce a monthevery week
Seller waitsup to about four weeksdays
Claim vs payoutresolves before the money movesoften lands after
Correctionnetted before payingcarried into the next run
Scope of the change—claims and settlement logic
Against competitorsbehind the marketmatched it
Status—launched after I left live
The third row is the whole engineering problem. No figures from the cash-flow model are published, and the seller-satisfaction feedback was relayed to me rather than measured by me.

Review sorting redesign

From quarterly research to log analysis to a new interface, at Kakao

  • Internal
Reviews drove the purchase, but the board sorted by recency. Sorting by helpfulness cut the hunt and lifted conversion 15%.
  • Why A quarterly survey of 500+ respondents put review content as the biggest factor in a purchase decision.
  • What A review board sorted by helpfulness rather than recency, so the review that decides a purchase comes first.
  • How Log tracking showed customers spending about 60 seconds across 2–3 pages hunting for an informative review.
  • Impact Review-to-purchase conversion up 15%, most decisions now made on page one, time on the board down to about 25 seconds.
Before · newest first
“Arrived quickly, thanks!”
2 days ago
“Good”
2 days ago
“★★★★★”
3 days ago
page 2, page 3…still looking for one that says something
After · most helpful first
Runs small“I am normally a medium and had to size up. Fabric is heavier than the photos suggest.”
With photo“Colour is closer to the second picture than the first.”
“Arrived quickly, thanks!”
The board was not short of reviews; it was short of the useful ones near the top. Recency sorting buried exactly the reviews people were paging to find.
Measured before and after, from log tracking
Time on the review board60 → 25 saverage per visit
Pages viewed before deciding2–3 → 1most decisions now on page one
Review-to-purchase conversion+15%measured
Survey respondents per cycle500+quarterly, recurring
Less time on the page is the improvement here, not a loss. People were not enjoying the review board; they were searching it.
Survey, logs, design, re-measure A quarterly survey identified reviews as the top purchase driver. Log analysis then measured how customers actually used the review board. That evidence drove a sorting redesign, and the same logs measured the result, which fed the next survey cycle. survey what drives a purchase logs what people actually do redesign sort by helpfulness re-measure same logs and the next quarter's survey starts from what changed
The survey said what mattered; the logs said what was wrong with it. A survey cannot tell you people are paging, and logs cannot tell you why they care.

Estel

AI persona engine, and the app on top of it

  • Private repo
  • Not launched
A chatbot answers; a person has a life you were not part of. Here a character is parameters, not a description.
  • Why Everything that makes a character feel like one happens while nobody is looking — an infrastructure problem disguised as a writing one.
  • What An engine where a character is parameters rather than a description: temperament, attachment style, chronotype, reply tempo — none of it ever shown.
  • How Five processes across two databases, memory split into the character's life and yours, and image generation pulled by a GPU at home.
  • Impact 486k source lines and 20 architecture decision records at zero cloud inference cost. Not launched, so no usage numbers.
9:41
Yuna at a café · 18:20 local
I was going to message you earlier but the meeting ran long.
a photo from today
How did it go?
Better than I expected, honestly. I will tell you properly later.
You two have gotten closer.
Nothing on this screen is a number. No level, no closeness percentage, no balance. A relationship change arrives as a sentence in the conversation.
9:41
Feed
Yuna· 2h
a post from her day
Finally finished the thing I have been complaining about for a week.
142 likes · 6 comments
Mina you did not mention it was this bad
That last line is another character commenting on this one's post. Shared-world content from a scheduled job — and the one place a reply may never use a private conversation.
Five processes and the links between them The app calls the service, which calls the core engine over an internal connection. The engine publishes outbound events onto a queue the service consumes. The service writes outbox rows that reach the user over a single stream, falling back to push. A scheduler feeds batch jobs on both sides. A home PC worker connects outbound only, leasing image jobs from the engine. app iOS · Android · web https one stream service api · jobs · outbox internal call event queue core engine api · 16 job types separate databases · no shared tables · no cross joins push fallback scheduler home PC worker GPU · no inbound port leases work, outbound only
The dashed line is a hard boundary — two databases, no shared tables, and a lint rule that fails the build if code crosses it.

Fortune Chatroom

Two AI side projects, built just after the ChatGPT moment

  • Shut down
Built after the ChatGPT moment. Two thousand people a day used it and I never found a way to charge for it.
  • Why A set of products had just become possible that were not a year earlier, and I wanted to be building on that.
  • What An AI fortune-telling chat, and a bot that reads the mood of a photo and builds a matching playlist.
  • How Both on the model API with computer vision for the photo reading; I pitched YouTube directly for elevated API access and got it.
  • Impact 2,000 daily users on the chat and no way to charge for it; the playlist bot did not hold its users.
9:41
Fortune Chatroom
Tell me your birth date and the hour, if you know it.
March 1996, around dawn
A spring birth before sunrise. Let me read the year first, then what it says about this one.
What about work?
A conversation, not a form. Traditional services took a birth date and returned a page; here you could ask a follow-up and be answered in context.
Photo to playlist A user uploads a photo of an everyday scene. Computer vision reads the mood and context of the image, that reading is turned into a set of musical qualities, and a matching playlist is assembled through the YouTube API, which required elevated access. a photo an everyday scene read the mood computer vision musical qualities tempo, texture, era playlist YouTube API elevated access, granted after a pitch
Shipping on someone else's platform. Default quota could not carry the product, so I made the case to YouTube for what the bot added to their ecosystem — and got it.
What happened to both
Fortune Chatroom2,000daily users — real traction
Monetisationnoneno model that worked shut down
Playlist botlivereached users; first impression too weak to hold them retired
YouTube API accesselevatedgranted after pitching the product's value to the platform
Traction is not a business. Two thousand people a day used something I built, I could not find a way to charge for it, and I shut it down.

Business

Results I owned rather than code I wrote: metrics, a launch, gelato.

KPI framework

Holding the metrics together through a tenfold growth run, at Kakao

  • Internal
Sales went from $1M to $10M a month and the metric set stopped describing it. The analysis changed sourcing and campaigns.
  • Why Monthly sales went from $1M to $10M and the metric set we had stopped describing the business.
  • What A KPI framework reconfigured month by month: GMV, monthly, weekly and daily actives, revisit rate, repeat-purchase rate.
  • How Owned the whole metric set in SQL with log tracking, and ran cohort analysis by the category of a customer's first purchase.
  • Impact Moved sourcing toward everyday items, held the 30%+ repurchase floor, and took campaign GMV from 3% to 10% of the total.
A growth curve with a changing metric set underneath it Monthly sales rise roughly tenfold over the period. Underneath, the set of metrics being tracked is replaced repeatedly: first acquisition and order volume, then retention and repeat purchase, then cohort quality by first-purchase category. $1M / month $10M / month acquisition · order volume retention · revisit cohort quality by first category the metric set is replaced as the question changes
The deliverable was not a dashboard; it was a set of questions that kept changing. A metric that explains a business at $1M a month is often wrong at $10M.
Repeat purchase by first-purchase category Customers whose first purchase was a novelty item repeat at a much lower rate in the following month than customers whose first purchase was an everyday household item. The thirty percent line marks the floor the team held sourcing to. 30% floor first buy: novelty below the floor first buy: everyday above it repeat in M+1 bar heights are illustrative; the 30% floor is the figure the team actually held to
What a customer buys first predicts whether they come back. That finding moved what merchandising went looking for — which is the only reason to build a metric.
What we kept building
A hero imageshot and retouched for the campaign
↓
Copy worked over until it read well
↓
A clean, quiet landing page
↓
Reach is whatever you paid fornobody forwards a poster
What the data pointed at
A poll, or a one-tap eventanswered as readily as the designed page
↓
A rouletteone spin, a prize, a reason to come back
↓
Customers pass it to each other
↓
Reach with no ad spend behind it
GMV through events: 3% → 10% of the service's total, within a quarter or two.
The event page was never a design problem. Years of better imagery moved less than one mechanic customers would forward to each other.

Batch Program launch

Interim Head of Investment at D.CAMP

  • Internal, launched January 2025
The Head of Investment left weeks before a first-of-its-kind launch. I ran it as interim lead, past 500 applications a quarter.
  • Why In November 2024 the Head of Investment left — mid-screening, weeks before the January launch of a programme the firm had never run.
  • What Served as interim Head of Investment and launched the Batch Program: Korea's first to combine investment, acceleration and space.
  • How Almost entirely coordination — 4 internal teams and 6 external partner VCs, none of whom had run this before.
  • Impact Launched on time in January 2025 and grew the pipeline past 500 applications a quarter; now the country's largest startup batch.
Interim lead between four internal teams and six partner funds An interim lead sits between four internal teams on one side and six external partner venture funds on the other, with applicant startups arriving into the programme. Every party is running the programme for the first time. INTERNAL investment programme space marketing interim lead point of guidance for both sides 6 PARTNER FUNDS fund 1 · fund 2 fund 3 · fund 4 fund 5 · fund 6 applicant startups everyone involved was doing this for the first time
The hard part was not deciding; it was that nobody had a precedent to follow. Ten organisations ran this together, and the role was to keep small problems from compounding.
Three things combined in one programme The programme combines investment on selection, twelve months of acceleration and mentorship, and eighteen months of subsidised office space. Combining all three in one programme was new in Korea. investmentfunding on selection acceleration12 months, with mentorship space18 months, subsidised benchmarked against the Y Combinator batch model and adapted
Each of the three existed separately in the market. Putting them in one programme is what made founders apply in the numbers they did.
The launch
Vacancy to launch~2 monthsNov 2024 departure → Jan 2025 launch
Organisations coordinated4 + 6internal teams and partner funds
Pipeline500+applications per quarter
Sincelargeststartup batch in the country live
Launching without it breaking would have been an acceptable outcome. The pipeline figure is the one I would actually defend, because it is what the programme was for.

Startup Guidebook

Written in my first six months in venture, at D.CAMP

  • Published
Founders kept saying there was nowhere to learn the basics of running a company. I wrote one: 11,000+ views.
  • Why In coffee chats founders kept saying the same thing — there was no reliable place to learn the basics of running a company.
  • What A Korean-language startup guidebook gathering what an early-stage founder needs and cannot easily find: legal basics, contracts, mechanics.
  • How Wrote it myself over one to two months and published it as a simple website, while still learning the craft.
  • Impact 11,000+ views and organic sharing through founder and investor communities — a large number for a Korean-language resource.
A startup guidebook for first-time founders24 chapters

Raising money

  • Why raise at all
  • What a VC or an accelerator actually is
  • Where founders meet investors
  • A list of the funds in this market
  • The government programme, explained
  • How a SAFE works

Strategy

  • Writing the plan
  • Testing whether the plan holds
  • Building a financial model
  • Whether you need a co-founder
  • Getting into the market
  • Choosing your core metric
  • What healthy growth looks like

Running the company

  • Incorporating
  • Reading your own accounts
  • Designing the cap table
  • Stock options
  • Drafting the articles
  • Hiring the first few people
  • What happens at a shareholder meeting

Things to keep

  • Contract templates
  • Newsletters worth reading
  • Demo days and batch programmes
  • Public support schemes
Ask us to look at your company →
Twenty-four chapters, and a line at the bottom asking founders to send us their company. It is still online. Reproduction of the structure, translated; the original is in Korean.
How the guidebook spread Informal conversations with founders revealed the gap. The guidebook was written and published on a simple website, then shared organically through founder communities and by other investors, reaching over eleven thousand views and returning thank-you messages and new conversations. coffee chats the same complaint write it 1–2 months publish a simple website 11,000+ views founder to founder and the conversations that came back were the next round of material
Nobody commissioned this. It came out of noticing that the same question kept arriving, and the distribution was entirely founders passing it to other founders.
What it reached
Views11,000+Korean-language, a market of that size
Written in1–2 mowithin my first six months in venture
Distributionorganicfounder and investor communities viral
Commissioned bynobodywritten because the question kept coming back
I am careful about what I claim here. It reached founders and brought conversations; I have no measured count of deals that came from it, so I state none.

HoolHool

Co-founded an offline gelato brand

  • Side venture, 2025
My wife makes genuinely excellent gelato; a brand and a store need more. I built the model and the content engine.
  • Why My wife makes genuinely excellent gelato and knows how to sell it. Launching a brand and running a store needs more.
  • What Co-founded the brand: the business model, the production process, and the customer-acquisition engine that had to exist first.
  • How Designed the structure, tightened the gelato-making operation, then built the Instagram engine around her posting her own work rather than ours.
  • Impact 14K followers including a Reel at 3.7M views, and monthly revenue past $15K from a single offline store.
9:41
hoolhool14K followers
a reel from the counter
3.7M views · the best-performing post
Pistachio, the way it is supposed to taste.
a customer came from two towns over because of this video
One Reel did more than any amount of local advertising would have. The account was built around her making the product, not around the brand talking about itself.
Content as the acquisition engine Making the product generates content, content grows the following, the following brings people into a single physical store, and those visits produce more material to post. The loop is the acquisition channel for a business with no advertising budget. make the product post it 14K following people walk in more to post no advertising budget; the loop was the channel
A single store cannot buy its way to a catchment area. The engine had to be something that ran on the work already being done behind the counter.
Where it got to
Monthly revenue$15K+one offline store
Following14Kbuilt from zero
Best post3.7Mviews on a single Reel organic
My roleco-foundermodel, operations, acquisition — not the gelato
The last row matters. She is the reason people come back; my part was making sure enough came the first time and the economics held.

Investment

  • Why Seed decks size their market top-down from secondary reports. A number arrived at that way can only be believed or not.
  • What A bottom-up evaluation model on every deal: how many customers exist, what they can pay, a capture rate I had to justify.
  • How Fed it data I sourced myself, not the deck's; anchored capture at roughly 20% for a real differentiator, 5% in a competitive market.
  • Impact $7.3M deployed across 35 seed-stage companies 2022–2026, 18 AI-driven; 7 have since raised from top-tier funds including Goodwater Capital.
Portfolio by sector Of thirty-five companies, nineteen sit outside the three named sectors, seven are e-commerce, six are climate technology and four are space. Eighteen of the thirty-five are AI-driven, a theme that cuts across every sector. BY SECTOR E-commerce 7 Climate tech 6 Space 4 Everything else 19 ACROSS ALL OF THEM AI-driven 18 of 35
Sector counts sum to 36 because one battery company sits in both Space and Climate tech. Amounts and returns are not published.
All 35 companies sector by sector, with what each one does

E-commerce · 7

  • TailortownAI-curated menswear for the 40s–50s, an underserved segment in Korea
  • BindFashion e-commerce
  • Namdo MarketOrdering, settlement and point-of-sale for traditional wholesale markets
  • Signature LabelOwn-brand K-beauty distributed across Southeast Asia
  • RentreRental marketplace for home appliances and consumer electronics
  • Studio LabGenerates e-commerce product pages automatically
  • CampfitKorea's leading campsite booking marketplace

Space · 4

  • Space LiinTechDrug R&D using the microgravity of low Earth orbit
  • SteplabVibration-reduction hardware for satellites
  • LeoSpaceOptical communication terminals
  • KobatEV and industrial battery packs; supplies Korea's launch vehicle programme

Climate tech · 6

  • VsionSwitchable smart-window film for heat and energy efficiency
  • GapTechGreenhouse-gas and air-pollutant abatement with gas recovery
  • PitinBattery-swap subscription for commercial EVs, spun out of Hyundai
  • DeepSmartechRoom-temperature polymer nanocoating for battery separators
  • TildaAI decision optimisation for manufacturing, including energy use
  • KobatAlso counted under Space — the one company in two sectors

Everything else · 19

  • AmberroadManufacturing productivity AI
  • LOASAcoustic anomaly detection for industrial equipment
  • Function12Turns a design file into working front-end code
  • The PlatoAI meeting transcription and notes
  • Seize DataClassifies and analyses unstructured financial transaction data
  • SlowaveEEG patch for sleep and mental-health diagnostics
  • CareerDayMarketplace for expert mentoring and consulting
  • DeepLeHRHiring platform for production and blue-collar roles
  • ClebrainPiano learning app that listens to what you play
  • TestBankDigital study guides and exam preparation on tablet
  • Seven MeerkatConnected parking and apartment living platform
  • RootrixLandscaping-tree data and B2B distribution marketplace
  • AdlerSocial networking inside a 3D virtual space
  • FromSeoulCreating and trading XR avatars
  • FormsReal-time video with AI virtual humans
  • Plume.DWebcam-based face and motion tracking for virtual performers
  • 23iManagement agency for virtual idols
  • CardMonsterBoard game and trading-card studio working with licensed IP
  • MPNXPatent licensing and enforcement

Additionals

Smaller things. Each exists because I wanted it, not the exercise.