Building an AI Startup · Ashaya Sharma
Elevate Venture

Building an AI Startup

Lessons from building Honeycomb AI in a niche corner of the market.

Ashaya SharmaCo-Founder, CTO, Engineer
02 — Credibility

Why should you listen to me?

Scar tissue.

The short version

“I’ve made all the mistakes in the book.”

Hired too early. Scaled before fit. Built features instead of outcomes. Now I help teams skip the painful middle.

03
03 — Case study

Honeycomb AI

Building a niche product in an AI space, where the moat isn’t the model, it’s the data nobody else can be bothered to collect.

Own modelsFrontier-augmentedHard-to-reach data
Lesson 01

Built our own models

In a niche, off-the-shelf models give off-the-shelf answers. Owning the model meant we could tune for the vocabulary, edge cases, and ground truth of our domain.

Own it
Lesson 02

Augmented by the rising tide

Frontier models keep getting better for free. We layered them on top of our own, using each for what it does best instead of betting the company on one.

Compose
Lesson 03

Chase the hardest-to-reach data

Anyone can scrape the easy stuff. The real moat was the data competitors gave up on: messy, gated, expensive to gather. That’s what made the product defensible.

Moat
03 — Track record

Products I’ve scaled

Closed deals with
DoorDashnoonSodexoEarls
Scaled to
2M+end users
Built & advised
1company sold
10+startups helped
04
04 — The new game

What separates the winners from the losers in this era?

When everyone can ship, the bar moves. Three things still decide who wins.

DistributionHard thingsWedge
Edge 01

Distribution

The best product no one hears about loses to a worse one with reach. Build the audience, the channel, and the wedge into how it spreads, from day one, not after launch.

Reach
Edge 02

Doing the hard thing

Anyone can vibe code a simple solution now. Not everyone is willing to build something genuinely difficult: the messy data, the unsexy infra, the problem that takes a year to crack. That’s where the moat lives.

Moat
Edge 03

Nail the wedge

One painful problem for one specific user. Win it convincingly before you widen the surface area.

Focus
05
05 — Tooling

What VCs look for in an AI startup

Three things investors actually pattern-match on when they meet an AI team today.

Data moatTalentDistribution
Signal 01

Unique data moat

Not just “we have data”: how hard was it to get, and how impossible is it for someone else to replicate? Easy data is table stakes; unique data is the whole pitch.

Moat
Signal 02

Talent on the team

Everyone is getting laid off. Top AI engineers are not. VCs underwrite the team first: who you’ve hired, who you can hire, and why they chose you over a FAANG offer.

Team
Signal 03

Ability to distribute

Early traction metrics and a credible path to $1M ARR fast. Can you get the product in front of the right buyers and close them, without burning a round on ads?

GTM
06 — The funnel

A tale of 30 startups

Most don’t make it. The one that does isn’t lucky. It’s built differently.

Start30
30 companies set out
No PMF12
−1818 drop out, never found product-market fit
Founder fallout2
−1010 drop out from co-founder conflict or chased other opportunities
Acquihire1
−11 gets quietly acquihired
Still going1
1 is still building
The one that survived

What they had in common

  1. 01
    Do not die
    Optionality compounds. Stay in the game.
  2. 02
    Unique technical expertise
    Something the next team can’t replicate in a weekend.
  3. 03
    Hyper customer-focused
    Closer to users than anyone else in the category.
06
06 — Velocity

Vibe coding + AI

Use AI to compress the build loop. Idea to clickable prototype in an afternoon, then learn, kill, or double down by Friday.

Rapid iterateRapid buildShip
Why it matters

10× faster prototypes

An afternoon of vibe coding produces what used to take a sprint. The cost of being wrong drops to nothing.

Build
How to use it

Build, then file the ticket

Stop arguing in spec docs. Ship a working demo, watch users react, then write the ticket for what survives.

Iterate
07
07 — Hard lessons

Mistakes worth stealing

The expensive ones I’ve already paid for. Take them. Skip the bill.

PMFHiringOutcomes
Mistake 01

Scaling before product-market fit

Ops

Costly
Mistake 02

Hiring ahead of clarity

A bigger team multiplies confusion before it multiplies output. Hire when the role is obvious, not when it’s tempting.

Painful
Mistake 03

Features instead of outcomes

Customers buy a job done, not a roadmap. Measure what changed for them, not what shipped from you.

Avoid
08
08 — Operating principles

What I run by

Three principles I keep coming back to when teams are stuck choosing between five good options.

ListenBuildHard things
Principle 01

Listen to customers

Principle 02

Build what they want

Principle 03

Do the hard things

07 — Diligence

Technical analysis

Four questions I’d stress-test any AI startup against before writing a check, or joining the team.

01
CAGR & bottom-up TAM
Is the market actually growing?
Forget top-down slide-ware. What is the true customer count from a bottom-up TAM: segments, ACVs, and a defensible growth rate?
02
Uniqueness of data / solution
How hard is this to rebuild?
Estimated engineering cost + token cost to recreate the ultimate solution. If a competitor can do it in a quarter, it isn't a moat.
03
Distribution tactic
Does the math actually work?
LTV > CAC for most startups. A clear channel, a repeatable motion, and unit economics that survive contact with paid acquisition.
04
Market-rate pay of the team
What could they get on the open market?
If the founding team could clear $500k+ at a frontier lab, the opportunity cost is the signal. Talent density compounds; mediocre teams don't.
08 — Final takeaway

Final takeaway

What investors are actually evaluating shifts at every stage. Build for the question you’ll be asked next, not the one you just answered.

Seed
Founders only.
There is no product to underwrite, no metrics to model. The bet is the people: their judgment, their unfair insight, their refusal to quit. Everything else is a story.
Series A
What did they build? How much can it scale?
Now the artifact matters. Is the product genuinely hard? Does the architecture, the data, and the motion compound, or does growth break it?
Series B + Beyond
What is the proof?
Charts replace narrative. Revenue, retention, gross margin, payback. The story has to survive the spreadsheet.
10 — Stay in touch

Let’s keep talking.

QR code linking to X profile @ashayasharma
X / Twitter

@ashayasharma

Product notes, builder threads, and the occasional rant.

QR code linking to LinkedIn profile ashayas
LinkedIn

in/ashayas

Work history, talks, and proper introductions.

Questions

01 / 14
← → to navigate