Prospect Mining
An investment-development co-pilot that shows its work.
Designing the interface for Prospect Mining — a generative-AI tool that helps Arcadea's investing team find, understand, and reach acquisition targets, and lets a non-technical Managing Director correct the model in ten seconds when it's wrong.
Finding a needle, by hand, in a haystack of thousands.
Arcadea acquires vertical-market software companies. Its investing team spends enormous effort researching thousands of candidates to judge one thing — is this a genuine VSaaS acquisition target? — and then reaching the right people. That work was manual, slow, and hard to keep consistent.
The Arcadea AI Team's answer was a co-pilot, not an autopilot: a tool built on Azure Cognitive Services that does the routine gathering, scoring, and drafting, then hands the judgement back to a person. My job was the interface — the part the Managing Directors actually touch. It's also where my engineering past pays rent: designing for a probabilistic system means understanding what the model can and can't know.
And building the correction flow for when it's wrong. That framing set the whole design problem. A probabilistic system making high-stakes calls is only used if the experts on the other side can see its reasoning, trust the parts worth trusting, and overrule the rest without friction. The screen had to earn that trust.
Score the target
Predict how likely a company is to be an acquisition fit — surfaced as a plain-language probability, with the reasoning one click away.
Draft the outreach
Combine existing email templates with more nuanced, per-company messaging — drafted for review, never sent blind.
Centralise everything
Keep every learning inside one tool through two-way integrations with Salesforce, SalesLoft, and LinkedIn.
“A powerful tool — yet it'll only be as effective as its human counterpart.” — Arcadea AI Team, internal launch note
What I owned, and what fought back.
I designed the product interface end to end — the information architecture, the screen layouts, the tab system, the data-state language, and the human-in-the-loop correction flows. The AI models, pipeline, and integrations were the Arcadea AI Team's.
- Streamlit as the canvas. A data-app framework, not a design tool — I worked within its component set to make something that still felt considered.
- Dense, expert content. Four ML models, summaries, contacts, and deal signals all converge on one company view without overwhelming it.
- A skeptical, busy audience. Managing Directors, not analysts — every screen had to reward a ten-second skim and a five-minute deep read equally.
- Trust in a probabilistic system. The hardest constraint wasn't visual. It was doubt: the design had to make the model both legible and correctable.
Thirteen automated steps — and two moments that belong to a human.
The pipeline mirrors the research the team used to do by hand. Mapping it revealed the real design insight: the interface only has to be excellent at the two decision points where a person steps in. Everything else runs quietly underneath.
One screen that answers “who am I looking at?” before you ask.
The Company Profile is the home of the tool. I split it into a fixed identity rail and a working area of four tabs, so context never leaves the screen while you move between lenses.
The deal ladder reads at a glance.
Every company carries the same panel, so an MD builds muscle memory. I gave the data a consistent colour grammar rather than leaving raw text: emerald for a strong signal or a clean “no” early in the funnel, amber for the counts that measure momentum. The probability sits at the very top, in words — High, not 0.9134 — because a person is meant to judge it, not decode it.
Skim first. Drill only where you're curious.
The Overview leads with a plain summary, then folds the evidence into expandable sections. An MD reads two paragraphs and moves on — or opens the drawers and interrogates every claim.
Evidence on demand
- The key phrase and predicted industry lead, so the model's headline claim is visible instantly.
- The VSaaS Identification Analysis drawer explains why the tool scored the company the way it did — reasoning, not just a verdict.
- Arcadea GPT is a second tab: a chat, with five ready-made prompts to give MDs a running start and a free-text box for anything else.
Ground Truth turns every skeptical expert into a teacher.
This is the heart of the tool. Four models predict a company's category, industry, key phrase, and outreach angle. When one is wrong, the MD doesn't file a bug — they fix it inline, and the correction becomes training data.
Designed so fixing it is easier than ignoring it
- A star rating asks one honest question first: how accurate is this listing overall?
- Each wrong field has its own gentle prompt — “Industry incorrect? Fix it here” — with the right control already in place.
- A free-text box captures the why when a dropdown can't, so the model learns nuance, not just labels.
- Nothing feels like punishment. The tone is a request for help, because a corrected model is the whole point.
The design brief was really doubt — and the answer was to make correcting the AI feel like a favour, not a chore.
A co-pilot that lives inside the tools the team already uses.
Prospect Mining doesn't replace the stack — it plugs into it. Salesforce stays the source of truth; SalesLoft runs the outreach. My design work here was the seam: showing what flows where, so nothing feels like a black box.
Automation you can audit
- Salesforce remains the core CRM. Prospect Mining reads from it to start, and pushes scoring and summaries back — a genuine round trip.
- SalesLoft receives the AI-drafted cadences. Each day the MD reviews the queue, sends what's right, and polishes what isn't.
- Sent, read, and reply rates flow back in — closing the loop so the drafting gets sharper over time.
Salesforce and SalesLoft are third-party products Prospect Mining integrates with; those screens are the integration surface, not my UI design.
A quiet palette, so the data can speak.
Arcadea's own identity is spare — a black wordmark on white. I kept the product in that register and spent all the colour on meaning: state, signal, and the human touch.
Type does the hierarchy. A characterful serif for company names and section heads, a readable serif for the dense AI summaries, and a monospace for data, labels, and the deal ladder — so numbers and states always look like facts, distinct from prose.
Structure repeats on purpose. The same rail, the same four tabs, the same colour grammar on every company. Consistency is what lets an expert stop learning the tool and start using it — the highest compliment a working interface can earn.
Designing for a machine that's often, but not always, right.
Legibility beats accuracy
A 91% model an expert can't inspect gets ignored. Showing the reasoning — the VSaaS rationale, the source drawers — mattered more to adoption than the last points of precision.
Correction is a feature, not an error path
Ground Truth treats being wrong as ordinary and fixable. Framing correction as helping — not reporting a fault — is what turns a tool into a system that gets smarter.
Meet experts where they work
Living inside Salesforce and SalesLoft, in plain language and a familiar layout, did more for trust than any amount of visual polish could have.