S3E1, season 3, summer
The only engineer
Where it started
Akro Ventures is a capital advisory firm. When I joined in April 2026 it had no software at all, and its deal process ran on paper. I was the first technical hire, and the only engineer.
What I built in three months
- Two versions of the website, then the deal platform underneath the business.
- A public intake form for startups, plus CSV bulk upload with column-synonym mapping, so a spreadsheet's column names never have to match mine.
- Claude, with structured tool calls, turning each raw submission into a deal summary and a hot, warm or cold qualification.
- An approval flow for outreach. The model drafts founder emails. A person approves each one on a single-use link.
- An n8n pipeline that reads the news every morning and hands over posts ready to go. It saved about ten hours a week.
- Search work: metadata per route, structured data and an llms.txt. The firm went from near-zero visibility to being described correctly in Google and in AI answer engines.
The rules the model runs under
- It can't send. A wrong email to a founder costs a relationship, so sending belongs to a person.
- It can't invent facts. The firm's claims about itself live in application code. The model writes around them and never reaches them.
- Approvals can't be replayed. Every token works once.
- Row-level security lets an anonymous visitor submit a lead and never read one.
The bug I'd pick
Corporate email scanners open every link in an incoming email to check it for phishing. The approve links were plain GET requests. So the scanners were approving deals by themselves, and nobody had clicked anything.
Approval moved to an authenticated POST. Opening the link shows the draft. Only a deliberate confirmation sends it.
What happened
More than 200 startup leads went through the CRM. The firm's ₹20 crore client fundraising mandate came in on it.
Next.js, TypeScript, Supabase, PostgreSQL, Claude API, n8n, Vercel
Next episode, S4E1
I asked my rejections what they had in common