Mts On Device Ai Ml

Member of Technical Staff (MTS) — On‑Device ML (MTS‑EDGE) — Kharagpur/Pune/Bombay (On-site Lab=Garage Work from Office Only) Put intelligence where it earns trust: on the device, by default. If you get a thrill from shaving milliseconds, shrinking models, and making voice feel instant without shipping data to strangers, this is your stage. Why this role exists Kai and Nav are only as trusted as their privacy and responsiveness. On‑device inference gives us both. Your work delivers real‑time speech, local understanding, and smooth synthesis with battery and memory budgets that respect the user—and the product. What you’ll do - Quantize, distill, and compile models (VAD/ASR/NLU/TTS/Small‑LLMs) for CoreML/Metal with streaming pipelines. - Build offline fallbacks and clever on‑device first, cloud‑when‑necessary routing with clear consent gates. - Instrument latency, accuracy, power, memory, and binary size; land targets and keep them there. - Partner with iOS to ship low‑friction voice UX (barge‑in, barge‑out, interruption repair). - Collaborate with Orchestration to design tool contracts that exploit edge capabilities (wake‑words, local context, cache). - Maintain a reproducible toolchain (PyTorch → ONNX → CoreML) with deterministic builds and golden benchmarks. 30 / 60 / 90 outcomes - 30 days: Baseline ASR+TTS pipeline running locally with streaming; p95 end‑to‑end voice turnaround 800 ms on a recent iPhone; battery/thermal tracingin place. - 60 days: Add wake‑word, VAD, and small‑footprint NLU; offline short‑task mode(notes, reminders, follow‑ups); accuracy/latency dashboards shipped. - 90 days: > 85% task‑success on the on‑device golden set; stable MOS for TTS; memory 300 MB at p95; 3% battery per10‑minute assisted session. Signals we’re looking for - You’ve shipped CoreML/Metal inference with real‑time audio and can show profiling traces. - You’ve done quantization (INT8/INT4), distillation, or sparsity and can explain the trade‑offs. - Comfort across Swift/C++ bridging, AVAudioEngine, and low‑latency I/O. - Taste for graceful degradation and offline safety. The stack you’ll touch CoreML, Metal, Accelerate, AVAudioEngine, Swift/C++, PyTorch, ONNX, Python toolchains, Instruments, power/thermal profilers. How we work Campus pilots, weekly ships, privacy by architecture, ruthless measurement, written decisions, small senior team. Compensation ₹1.0L – ₹1.5L per month + meaningful equity. Board‑approved bands; refreshers tied to shipped impact. Apply Email talent@hushh.ai with subject “MTS‑EDGE — Full-time / Internship” Include code links and five bullets on how you beat a latency/accuracy/power target in the real world with clear examples and Github repo references with the best project contributions you've completed so far in your personal and professional life / portfolio. Please note we are a small team with minimal overhead so please be patient with our responses as we have received over 100,000 applications for every job we open. We will prioritize candidates with completed applications and the best projects that showcase their open-source and community work along with their individual contributions.