All case studies
FinTech · AI Shipped Jan 2026 · 28-day build

A wealth concierge that drafts in seconds, not days.

Havéne is a digital wealth advisory firm serving high-net-worth millennials. Their advisors were drowning in portfolio summary work. We built an AI agent that drafts every summary, performance report and client memo — at a quality their senior advisors approve in one pass.

The Challenge

Each advisor was spending 12+ hours/week on client-facing writeups — quarterly summaries, performance attribution, market commentary. The work was high-stakes (compliance-sensitive) and high-volume. It was also the work that paid the least.

What we shipped

  • Portfolio summary generator (per-client tone-tuned)
  • Performance attribution AI
  • Compliance pre-flight checks (FINRA-aware)
  • One-click market commentary briefs
  • Advisor approval workflow

Stack

  • Python · FastAPI
  • Claude Opus 4 · GPT-4o for evals
  • Custom RAG over market data + prior writeups
  • Postgres · Redis · S3
  • AWS · Terraform
32hSaved / advisor / week
94%First-draft approval rate
28dFrom kickoff to live
0Compliance flags in 6 mo

Why this was the hardest project we've shipped

Finance is unforgiving. A wealth summary with a hallucinated number isn't a typo — it's a compliance event. We approached this knowing that "feels right" wasn't good enough; the system had to be verifiable.

"Our advisors were skeptical. The first draft I read, I tried to find something wrong with it. I couldn't. That's when I knew." — Priya N., Chief Investment Officer, Havéne

The trick: never let the model invent a number

Every numeric claim in every draft is grounded in a structured source. The agent doesn't generate performance figures — it retrieves them, then writes prose around them. If a number isn't in the structured data, the agent writes "[need data: 12-month return]" instead of hallucinating.

This single architectural decision is why first-draft approval rate hit 94% — advisors trust the numbers because the numbers came from their own data, not the model.

Compliance as a feature, not an afterthought

Every draft passes through a compliance pre-flight layer trained on FINRA's marketing rule (Rule 2210) and Havéne's internal style guide. Flagged drafts are sent back with the offending phrase highlighted. After six months in production, zero compliance flags have made it to a client.

What we'd do differently

We'd build the eval suite before the agent. Half of the first 28 days was spent inventing ways to measure "is this draft good?" — questions a finance domain expert could have helped us specify upfront. On the next finance project, we did exactly that.

"This gave us a third of our week back — and the part of the week we hated. Our advisors now spend that time on actual client relationships."

— Priya N.CIO, Havéne
Next case study →

Vigor Beans — a D2C coffee brand running on cruise control

Read next case