Economics — the numbers that decide ₹250

Status: template — filled from Day 3 measurements. The single most important number in the POC: measured ₹ inference per bot per month at the defined usage profile.

Usage profile (the yardstick)

One bot = 30 chat messages/day + 1 routine/day (e.g. the 6am inbox digest). Messages average ~4k input tokens (system prompt + memory recall + history) and ~400 output tokens. A routine averages ~6k input / ~800 output.

Measurements (fill Day 3)

Metric	Target	Measured
Bots live on the 8GB slice	≥10	—
Idle RAM per bot	≤250MB	—
Platform overhead (manager+gateway+telegram+proxy+search)	≤1.5GB	—
p95 first-token latency (chat)	<4s	—
12h soak: routines fired on schedule	100%	—
Inference cost per bot/month (profile above)	≤₹60	—

Cost model

Per 16GB node / month:
  hardware amortized (₹35k / 36mo)            ≈ ₹972
  electricity (~15W avg, ₹8/kWh)              ≈ ₹90–150
  ISP share (host's existing line)            ≈ ₹0–1,800
  host payout                                 ≈ ₹400
  total                                       ≈ ₹1,550–3,350

Per bot:
  infra = node total ÷ bots_per_node
  inference = measured from gateway spend log (₹)
  margin = price − infra − inference

Verdict table (to complete)

Price \ bots/node	10	15	20
₹250	—	—	—
₹299	—	—	—

Notes on levers observed during the POC (fill in): free-tier routing for routines, cache-hit rates, peak/off-peak timing of scheduled runs, model mix per persona.