Finance, software and AI
Rishabb Shankar
I build financial models and the software around them. I like being able to see why a number came out the way it did.
Financial analysis
United Airlines and Apple
I start by asking which assumptions actually move the valuation.
The United Airlines DCF puts the base case near $121 a share against a $105 price on Aug. 29, 2025: a HOLD. In the Apple study, the Services mix explains 172 of 290 bps of margin gain. Both are built from filings.
Software
Spreadline
Spreadline takes synthetic financial inputs, works out coverage ratios, checks them against lending rules and writes a short memo explaining the call. It’s a prototype, built so you can trace every number back to where it came from.
When something breaks, I want the system to have left enough behind to tell me why.
AI systems
Agents for my agency
The marketing agency I built runs on a team of agents.
None of them can post, send a message or spend money without me. The one that reads customer messages only sorts them. I write the replies myself, because they should come from a person.
View the system map
System map
- BootEach session starts by recounting jobs and reviews straight from the log, so no agent works off an old summary.
- RouteOne orchestrator reads the brief, works out what phase the business is in and passes the job to whichever specialist owns that area.
- RosterEach specialist owns one area. Two sit dormant until the business reaches the phase that needs them.
- reviews
- local search
- social
- decks & carousels
- booking flow
- brand
- measurement
- partnerships
- inbound triage
- paid ads (dormant)
- email & SMS (dormant)
- Brand gateAnything going out passes a brand review first, and every caption gets linted before a person sees it.
- ChecksA repository audit and a booking-log check fail loudly when something is off, so a stale fact stops the work before it spreads.
- OperatorPosting, sending and spending all stop here. The system drafts, then waits for me.
It runs on a client’s real bookings and customer messages, so the code stays private.
Language models
Izumi
Izumi is a chat workspace that uses several language models under one versioned set of rules. For every reply it keeps track of which prompt produced it, which model answered and why.
It’s still a private project and not ready for the public, so it’s invite-only for now.