Ram

Every case on the portal had to be opened, sent back, picked up, pinned on the map, given its document, filled, submitted and finally sent to the client. By hand that is nine screens per case, a few hundred times over.

The runner reads the case list from Excel and does it the way a careful person would. Before every click it reads what the portal is actually showing, waits between requests, and if you press Stop or the connection drops, it carries on from the exact step where it paused.

• Its requests match the browser’s own byte for byte, checked by golden tests.
• 37 tests run against a mock portal, so the real one is never touched while testing.
• A case with a missing value is held back with the reason shown. Nothing is guessed.

When a job is too big for one prompt, I don’t write a bigger prompt. I split the work into small packets, run agents on them in parallel, and give every agent the same written brief.

The brief is the real work: what done looks like, read-only access by default, no copying a value from one item to another, and a blank with a written reason whenever something can’t be proved. Every agent checks its own output before handing back, then plain code sweeps the merged result: ranges, totals and duplicate rows. A person signs off at the end.

Every property valuation has to quote the government’s minimum land rate for its area. The official portal only searches by name, and people in the field come back with a coordinate.

I wrote a harvester that keeps a local copy of every rate for all 41 districts, with history and the boundary of each colony. A pin now returns its rate in milliseconds without internet, plus live land-use, zone and khasra from the state GIS, and makes the map images that go with it. It is the first tool in the valuation industry that gives an up-to-date DLC rate straight from a coordinate.

Some portals want the same data typed in field by field: 60 to 380 fields a case. The extension reads the case’s Excel file and fills the form. It finds values by their labels rather than fixed cells, so a slightly different layout still works, and it flags a section in red instead of guessing. Saving and submitting stay a person’s click.

It began as five separate extensions. I merged them into one core with a small plug-in per portal after a review found the same bugs copied into all five.

36,000 case records sat in separate monthly spreadsheets. I loaded them into SQLite and built a dashboard on top: which clients are growing, which have gone quiet, how the month is tracking, which weekdays are busiest.

The alerts read like a colleague wrote them: “no cases from this client for 74 days, worth a call.”

Auction notices are spread over three websites in three formats. This collects them daily, keeps Rajasthan only, pins each property and compares its reserve price per sq ft with the median for its district, so a cheap listing stands out.