Case 02 · Venture Pilot · Human Design
Cutting deal screening from 4 days to 6 hours.
0→1 agentic workflow and interface design for a $163M pre-seed portfolio, 102 companies — sole designer and front-end developer, one quarter.
Screening time and trust from usability testing; capital measured over the two quarters after launch.
Four days of research, done twice.
The firm matches pre-seed startups to each investor's thesis. An advisor did it by hand, but investors didn't trust a shortlist they couldn't interrogate — so they re-ran the research themselves. Four days of work, done twice. An earlier fix, feeding the data to an LLM, just returned a wall of text.
This was never an AI problem — the AI was already there and already failing. The gap was structure. You don't make a hard call faster by compressing it to a score; you make it faster by structuring the reasoning so it's legible at a glance. Structure, not compression, became the spine of everything that followed.
The starting point — the firm first fix
The journey — four stages
- 01Calibrate
Set thesis, sectors, risk tolerance.
- 02Discovery
Scan the deal flow — reasoning inline, cited.
- 03Monitor
Signal-first pipeline; risks alongside upside.
- 04Finalising
Deep-dive with calibrated friction on high-risk calls.
What I owned on an 8-week clock.
Solo design and front-end, team of four (CEO, CTO, a backend AI engineer, me), no design layer above me. I owned how the score is presented, not the rubric behind it.
Ownership
Owned end to end
Agentic workflow architecture · Investor-facing interface & full UI · Score presentation and disclosure hierarchy · Front-end build · Research synthesis
Ownership
Shared
Direction & scope (with CEO/CTO) · Interaction details, esp. the keyboard system (with an engineer)
Ownership
Handed off
Backend engineering handoff · QA
Ownership
Not owned
Rubric criteria & scoring logic (backend AI engineer / CTO) · Backend pipelines · Model tuning
The conditions that shaped every call.
Trust was the real constraint. Every screen had to earn confidence.
Constraint
Three months, 0-to-1, solo
A quarter to a deployed product from nothing, one person designing and building it. No time to be precious.
Constraint
Poor, self-reported source data
Pitch decks are self-reported, inconsistent, often incomplete — exactly why the agent had to escalate instead of guess.
Constraint
Scope pressure was the real constraint
Stakeholders wanted more features; the product's value was removing them. Defending the scope was defending the outcome.
Four decisions carried the eight-week clock.
Four decisions carried the quarter: whether to fit the existing workflow or replace it, where the AI's reasoning had to live, how to make the recommendation honest enough to trust, and where to slow the user down on purpose.
Fit the workflow, don't replace it
Adoption over reinvention
The signal
A lift-and-shift would force investors to learn a new process — right as an entire party, the advisor, was being removed from the middle of it. Nothing kills adoption faster than an unfamiliar tool.
Options considered
Why rejected
A new process trades the whole point — adoption — for novelty. A bolt-on leaves the duplicated research untouched.
Why chosen
The calibrate → discover → monitor → finalise flow stayed familiar; underneath, the agent replaced the advisor's manual matching and the investor's second research round. Adoption over reinvention.
Verdict
The investor's experience barely changed. Everything behind it did.
Put the 'why' where the decision happens
Usability testing · trust as constraint
The signal
Investors distrusted opaque scores and dug too long to build confidence. Explainability couldn't be tucked in a hover.
Options considered
Why rejected
A hides trust in an edge case. B forces a context-switch that kills the speed we were solving for.
Why chosen
Every recommendation carries its reasoning inline — ‘High-velocity team, shipped v1 and v2 in 3 weeks, strong developer-community signal’ — each claim cited, with Scout revealing more on demand. Denser cards, but an AI that shows its work.
Verdict
That's the move that turned the AI from a black box into a decision partner.
An AI that doesn't guess — it escalates
Escalation over fabrication · signal + code-push
The signal
The most dangerous thing a confident agent can do is answer a question it has no data for. And a recommender that only surfaces positives reads as a salesman, not an advisor.
Options considered
Why rejected
A cheerleader surface that only says ‘yes,’ or an agent that guesses past its data. Both read as marketing — and both are where trust breaks.
Why chosen
Risk dampens easy enthusiasm — that's the point. Balanced, sourced, and honest about its limits is what makes the green flags believable.
Verdict
Knowing when to escalate is what makes the answers worth trusting.
Friction, placed exactly where a costly mistake hides
Automation bias at high-confidence + high-risk
The signal
The most dangerous moment in a fast tool is a high-confidence score built on a thin reason — exactly when a user stops thinking and clicks. Automation bias is highest where the stakes are.
Options considered
Why rejected
Any pattern that let a high-confidence + high-risk fork pass without a pause — the moment automation bias does the most damage.
Why chosen
Everywhere else I was cutting friction; here I added it on purpose. A tool that only speeds you up will eventually speed you into a mistake.
Verdict
The one place I made the tool slower is the place it earns the most trust.
Craft decisions that carried the four beats.
Five supporting choices — one line each, evidence tagged where it exists. They keep the four beats readable without becoming beats of their own.
- 05.01
Bought the foundation, built only what was needed
A lean 30+ component shared library, one language across design and front-end — ~40% faster handoff.
- 05.02
Predictable information zones
A consistent card grid trains the eye to scan the same way every time.
- 05.03
Contextual triage
Overview / Team / Technical / Market / Financials tabs let investors go as deep as the decision needs — no deeper.
- 05.04
Above-the-fold prioritization
TESTINGUsers engaged most with high-scoring, above-the-fold companies, so the strongest signal leads.
- 05.05
Deferred loading
Secondary detail sits behind ‘Load more’ to keep the first scan fast.
- 05.06
Keyboard system
RESEARCHAnalysts live on shortcuts — built with an engineer for power-user speed and motor accessibility.
What changed in the numbers, and in a sentence.
Screening — thesis to a shortlist and a booked meeting — dropped from about four working days (~32 hours) to six, roughly an 81% cut. In testing, 13 of 15 investors decided without second-guessing — and, tellingly, stopped running their own second research round and booked meetings directly. The survey said they felt confident; the behavior proved it.
The advisor wasn't removed — repositioned, handling only the cases the agent escalates. Faster screening plus that reclaimed capacity is where the 15% more capital deployed came from. On the build side: a 30+ component shared library, ~40% faster handoff, 15+ screens in the quarter.
Measured outcomes
Screening dropped from four days to six hours.
Screening time and trust came from usability testing; the 15% lift in capital deployed was measured over the two quarters after launch.
4d → 6h
Deal screening time
+15%
Capital deployed
13 / 15
Decided without second-guessing · usability test
“I re-run four days of research the firm already did, because I can't tell why the AI picked these.”
“I used to second-guess the AI. Now it feels like a teammate.”
What Venture Pilot taught me about designing AI.
- P.01
Speed comes from structure, not compression.
You don't make a hard decision faster by hiding information — you make it faster by organizing it so the reasoning is legible at a glance.
- P.02
Trust is available depth, not removed depth.
Investors skipped their own research because they could verify the summary on demand — and because the agent escalates instead of guessing. You earn trust by making depth available, not by hiding it.
- P.03
Good UX adapts to expertise; it doesn't abstract it away.
The keyboard system, the depth-on-demand tabs, the calibrated friction — all of it meets experts where they are instead of dumbing the tool down.
Next, I'd wire in selective external signals — market, funding, and regulatory events — to strengthen context without reopening the wall-of-text problem the product was built to solve.