FULL STACK · ASR + PUNCT + TRANSLATE + TTS + PHONEME | PHONE-CLASS SD 8 GEN 2 / A16 / TENSOR G3
The whole stack fits in your pocket.100% ON-DEVICE
The industry's playbook is simple: ship a thin app, send audio to a GPU server, charge per minute, hope nobody notices the latency or the bill. Pramiteeh's playbook is the opposite — ship the whole stack on the phone. Five models — streaming ASR, PunctFormer, Translate, TTS, phoneme — total under 450 MB, run on a mid-range Android in real-time, no internet, no telemetry, zero per-call cost. We are five engineers, self-funded. The numbers below are real.
Full-stack size
438MB
All 5 models combined. NLLB 1.3B alone: ~5.2 GB
Peak RAM
~620MB
Active inference, fp16+INT8. Sarvam Saaras: server-only
Real-time factor
0.18×
ASR on Snapdragon 8 Gen 2. Whisper-large-v3: 1.4×, server
Battery / hour
3.8%
Continuous live translate. Cloud equivalent: 11–14% (radio)
Cost / 1k calls
$0.00
Pay once, run forever. Cloud: $14–$28 per 1k
01
Model size, to scale
The bars below are drawn proportionally. Each Pramiteeh model is a sliver next to the cloud-class equivalents — and the entire Pramiteeh stack still fits inside a single one of them.
Code-switch (Hinglish, Tanglish)natural mid-sentence language flips
native
yes
word-level only
poor
Battery drain · 1 hour liveflagship Android, screen on
3.8%
~12% (radio + screen)
~13%
~14%
Funding required to shipcapital efficiency
$0 outside salaries
₹99 crore subsidy
Google scale
$10s of billions
03
One battery, one workday
Same phone — Pixel 8 Pro, 5,050 mAh, screen on at 50% brightness — running a continuous live-translate workload. Watch what happens over 8 simulated work-hours. The "Cloud Stack" version also drains the radio for constant audio upload.
NETWORK · ONLINE
SIM SPEED · 300×
Pramiteeh on-deviceASR + Translate + TTS · all local
A modest assumption: 1 million monthly active users, each making 200 short translation/transcription calls per month, ~1 minute each. That's 200 million calls and ~3.3 million minutes of audio per month.
Pramiteeh on-device
$0/mo
ASR compute · phone CPU/NPU$0
Translate compute · on-device$0
TTS compute · on-device$0
Audio bandwidth · stays local$0
Model distribution · CDN, one-time per user~$1.2k
Per user, per year
$0.014 · cents
Cloud-stack equivalent
$108k/mo
ASR API · ~$0.006/min × 3.3 M min$19,800
Translate API · ~$15/1M chars × 4 B chars$60,000
TTS API · ~$4/1M chars × 4 B chars$16,000
Audio egress · ~$0.09/GB × 145 TB$13,000
Per-call latency · network round-trip+ 280 ms
Per user, per year
$1.30 · 90× more
05
Why this matters
For every Indic AI use case that actually scales — a tutor in a Tier-3 town, a call-centre agent in Vijayawada, a farmer using a feature-class phone, a student in a school with 2G coverage — cloud is the wrong abstraction.
The audio shouldn't leave. The bill shouldn't compound. The model shouldn't fail when the radio is weak.
Pramiteeh ships the whole stack — ASR, punctuation, translation, TTS, phoneme — under 450 MB, on a mid-range Android, with no telemetry and no API call. We are five engineers building from our own salaries, and our models match the giants on benchmarks while costing nothing per call. That is the bet.