Turning Z47's edge into a system its whole team can use.

A firm's edge lived in people's heads - and every attempt to systematise it broke.

Every venture firm runs on an edge - the judgement that lets it see what others miss. At most firms that edge is scattered: some of it lives in a partner's head, some in old memos and deal folders, some in the one person who has seen a hundred companies in the same market. Z47 - formerly Matrix Partners India, an early backer of Razorpay, Ola, Dailyhunt and OfBusiness, with more than $3.5B under management and 10 unicorns - had that edge in abundance, and the same problem everyone has: it wasn't written down anywhere the whole team could use.

They had tried to fix it themselves, more than once, and each attempt broke. Change one thing and something else fell over - the logic had dependencies, and every edit really needed an engineer. A system only engineers can safely change is a system a partner will never adopt. That is exactly where most internal-AI efforts stall, and it is where Z47 was when they called.

Two artifacts sit on nearly every deal - and both were slow and uneven.

Produced by hand they are slow and uneven, and the quality depends on who did the work and how much time they had that week. The cost isn't one thin brief - it's every deal decision made on a shallower read than the firm is capable of, and judgement that walks out the door when a senior person does.

We built for change-safety first.

Speed on its own would only have rebuilt the same fragile thing faster. The hard part was the one Z47 had already hit: encoding a firm's thinking so it is both consistent and changeable. So we built for change-safety first.

• Front-loaded the risk. Weeks 1–2 went at the hardest parts - channel integration and the first end-to-end workflow - alongside the safety guardrails, before anyone relied on it.
• Made a non-engineer's edit safe. The team changes the system through chat, and each edit routes to the right place in the prompt despite the dependencies underneath - so behaviour changes without the system breaking, and without an engineering ticket. This is the exact thing that broke in-house.
• Kept the IP with Z47. The frameworks live in prompts and context Z47 owns; realfast built the governed system around them, not a black box to rent back.
• Secure by design. Allow-list access, per-user boundaries, and credentials never exposed to the agent.

Three things that decide whether this works.

• Adoption lives or dies on who can safely change the system. If only engineers can edit it, the people whose judgement you're encoding will never own it - build the edit path for them first.
• Meet the team where they already work. On WhatsApp and email, adoption is a message; a new interface is a project.
• Keep the judgement with the firm. The frameworks are the client's IP - the system should wrap them, not replace them or lock them in.

Z47 owned the judgement; we made carrying it lighter.

Z47 owned the thing that matters most: the investment judgement itself - the frameworks, what a good read looks like, and the call on whether an output was right. That accountability isn't something we could take off them, and shouldn't be. What we did was make carrying it lighter - the edit path lets a partner change the system's behaviour directly instead of briefing an engineer, and the guardrails mean they can do it without fear of breaking what already works. They validated it on real deals; we made sure validating it never meant trusting a black box.

“

Our edge - how Z47 thinks about a deal - lived in people's heads, and every time we tried to systematise it ourselves, it broke. Now anyone on the team can shape how we think, and nothing breaks. realfast is one of the very few teams that can actually get agents working in production, on complex workflows.