Drishti — दृष्टि

1 The whole system on one page

Everything lives on a single cloud VM we reach through Delta's internal access tool. The VM does all the thinking and writing; the public website is just a static page that reads pre-computed files.

The website holds no data of its own — it fetches the latest JSON straight from Cloudflare R2 (30-second cache). That's why a new signal appears without ever redeploying the site.

2 One cycle, every 15 minutes

The poller wakes about 96 times a day. Each run is three phases in order, then it sleeps to the next quarter-hour. This is the engineer's view of the four steps on the How It Works page.

If a data feed is down it degrades to n/a rather than failing the cycle, and a watchdog restarts the service if a cycle ever hangs.

3 Three loops, one database

It isn't only the 15-minute cycle. Two faster loops run alongside it so closes and trades happen in near real-time. All three write to the same DuckDB, so a single in-memory lock decides who writes when.

4 From a code change to live

Code is only ever edited on a Mac — the VM is a deploy target, not a workshop. Two paths reach production.

Logic / runtime changes

Website (shell) changes — this very tab

There's no CI gate today (org Actions are paused), so tests & linters run locally before every push. The ~96 publishes a day stay safely under Vercel's free-tier deploy limit because data goes to R2, not Vercel.

5 The stack at a glance

Every moving piece and the one job it does.

6 Words you'll hear

The terms that come up when we talk about how Drishti runs.

• PollerThe main loop that wakes every 15 minutes to generate new signals.
• Watcher (fast tracker)A faster 60-second loop that only checks open signals for a take-profit or stop, so closes are near real-time.
• ExecutorThe loop that places and protects real orders on Delta — off by default, gated behind explicit flags.
• PublisherPhase 3 of the cycle: writes the JSON files and uploads them to Cloudflare R2.
• Write-lockThe in-memory turnstile that lets only one loop write to DuckDB at a time.
• Cohort / prompt versionA labelled batch of signals from one prompt version, so results from different logic aren't mixed.
• systemd timerA scheduled job (like cron) for the daily & weekly extras — reviews, digests, macro and ETF data.