The fifth pillar of the stack. PramiteehLM is a compact language model that runs fully on-device — no cloud, no API, no per-call cost. The point isn't to compete with frontier LLMs; it's to give the rest of the Pramiteeh stack a local brain for the things that actually need one: short summaries, intent extraction, structured replies, code-switched responses. Runs today on a laptop. Android is the current work.
| Capability | Status | Use case |
|---|---|---|
| Instruction followingsummaries · extraction · classification | working | call summary in customer-support pipelines |
| Streaming generationtoken-by-token output | working | real-time response composition |
| Structured intent extractionJSON output | working | downstream of streaming ASR |
| Indic language reasoningHindi · Telugu · Tamil prompts | testing | code-switch prompts in customer flows |
| Android integrationrun on phone · zero cloud calls | in progress | full Pramiteeh stack on a single device |
| On-phone speed optimizationtuned for ARM | in progress | better tok/s on flagship Android |
| Lightweight on-device retrievallocal doc grounding | queued | on-device assistant memory |
| Per-domain adaptersswap behavior without re-shipping | queued | client-specific tuning |
A simulated session showing PramiteehLM responding on a laptop — the same flow being wired into Android now. Tokens stream at the real measured rate.
A small model on the phone won't write your novel. But for the things the rest of the Pramiteeh stack actually needs — short, fast, structured language work — it's the right shape.
After ASR produces a transcript, PramiteehLM compresses it to a 3-bullet summary + sentiment + next-action — never leaves the phone. Drop-in for customer-support pipelines.
From a customer utterance, emit JSON with intents, slots, and confidence. Pairs naturally with the streaming ASR demo — fan-out to retrieval happens here, not in the cloud.
Generate a reply that matches the user's register — Hinglish in, Hinglish out. PramiteehLM holds context locally so multi-turn doesn't pay a network round-trip on every step.
From the phoneme model's miscue output, generate the next coaching prompt — a minimal-pair drill, an encouragement line, the next sentence to read.