Kashish Dugar

Moonlight Build Log: Papa and I are building Moonlight with AI agents. I am the tech lead. I make the decisions and check the work. The agents write the code.

My paper graph

Today I drew the real data on paper before any code. PAH is the disease in the middle. Around it are five proteins: BMPR2, KCNK3, ACVRL1, EDNRA and PDE5A. Around each protein are its medicines.

• An arrow up means the medicine turns the protein up. An arrow down means it turns it down.
• A medicine and its tablet form go in one bubble, like Doxapram and Doxapram Hydrochloride.
• EDNRA has 16 medicines, so I made a tally grid for it.

Sorting the medicines

I ask two questions, in this order:

1. Is it approved for at least one disease? If not, cross it out and stop.
2. What is its link to our disease? No link is a new clue. A phase means it is being studied. Approval means it is already approved, so it is not a clue.

My numbers for the whole page:

• 16 new clues
• 2 being studied
• 6 already approved for our disease
• 12 crossed out
• 36 in total (40 rows on the sheet, but 4 pairs are one bubble each)

Claude counted too, and we got the same. So my drawing is now the answer key. The code has to get my numbers.

New words

• Inhibitor, antagonist, blocker: three ways of turning a protein down. An inhibitor is like a hand on a spinning wheel.
• Agonist, opener: two ways of turning it up. An agonist rings the protein’s doorbell.
• Phase 1, 2, 3: tests in people, from the first to the last. Approval comes after phase 3.

My errors

• I drew Doxapram in two bubbles first. My own rule says one.
• I missed Nebentan in my grid.
• I gave the Sitaxentan pair two ticks. A pair gets one, the stronger one.
• I put Sparsentan under “approved”. “Best stage anywhere” and “Link to our disease” are two different columns, and I mixed them up.
• I crossed out two PDE5A medicines that were approved.

If it is easy for me to mix up, it is easy for an agent. So these are now tests.

My new merge rule

Nothing gets merged until the agent explains it in five plain sentences and I answer questions about it. One agent said the filter used to be a bin and now it is a sorting tray with labelled slots.

My decisions today

• Medicines already approved for the disease are removed. They are not clues.
• WITHDRAWAL stays at the bottom of the stage ladder. A medicine taken off the shelves does not count as approved.
• If Moonlight meets an action word it does not know, it says “unknown”. It never guesses.

The commits

• 32372c8 Merge pull request #2: the saved data from Open Targets
• 970007d Ticket 1.3b: rulebook and logbook update
• af4dc72 Merge pull request #3
• 49369d2 Ticket 1.3: record types. 23 tests pass, and my six examples are in them.
• 9cb2f8b Merge pull request #4
• ec0e15d Add files via upload: my drawing, in the docs folder

Three agents are working tonight: the fetch module (pull request #6), housekeeping (pull request #5) and the filter.

Tomorrow

I review the fetch and the filter. The filter has to sort the medicines the same way I did on paper.