Senior /Staff Inference Engineer

Uniphore; Palo Alto (CA); On-site; USD 180,000 - 260,000; Full time

Job summary

Uniphore is building an enterprise AI platform where agents operate across voice, video, text, and structured data in real time. We seek a Senior or Staff Inference Engineer to own the inference layer and deliver a context-aware runtime serving finetuned SLMs across cloud, on-premises, and multiple GPU platforms.

This is a hands-on IC role collaborating with research, model engineering, knowledge systems, and platform teams to push performance, reliability, and portability across NVIDIA and AMD

Qualifications

• 5+ years engineering experience, including 3+ years with GPU systems or ML serving.
• Experience with production deployments of GPU-accelerated inference.
• Proficient in Python and C++, with ability to read CUDA/C++ kernels.

Responsibilities

• Own the inference layer across cloud, on-premises, and GPU platforms.
• Develop context-aware runtime, KV-cache, and prompt-prefix caching.
• Implement and optimize speculative decoding, attention kernels, and multi-GPU parallelism.
• Ensure production reliability with observability and safe rollout.

Skills

GPU systems

distributed inference

ML serving

Python

C++

CUDA

Tools

vLLM

SGLang

TensorRT-LLM

CUDA C++

HIP kernels

Job description

Uniphore is the Business AI company. Our sovereign, composable and secure AI platform connects enterprise data, fine-tunes AI models and deploys agentic AI across the enterprise. We empower every worker to boost productivity and help businesses grow faster, operate smarter and reduce costs. Trusted by more than 2,000 businesses globally, and recognized on the Deloitte Fast 500, Uniphore delivers on the promise of AI as a transformative force for business.

Job Description:

Staff Inference Engineer –

Uniphore is building an enterprise AI pla4orm where agents operate across voice, video, text, and structured data in real time. We are looking for a Senior or Staff Inference Engineer to own the inference layer — combining deep model inference expertise with systems engineering to deliver a context-aware runtime serving finetuned SLMs across cloud, on-premises, and widely used GPU platforms.

This is a hands-on individual contributor role working closely with Uniphore’s research, model engineering, knowledge systems, agent, and pla4orm teams.

What You Will Build

Context-aware runtime: Integrate context selection, compression, conversation compaction, and persistent task state with model execution. Adapt context budgets, model selection, and reasoning effort to each task.

KV-cache and computation optimization: Build paged KV-cache and prompt-prefix-cache management, cache-aware routing, retention, eviction, and offloading. Preserve correctness as context, permissions, model versions, and adapters change.

Inference performance: Implement and tune speculative decoding, attention kernels, continuous batching, quantization, chunked prefill, and multi-GPU parallel execution.

SLM and adapter lifecycle: Build model and adapter lifecycle management, multi-adapter serving, zero-downtime hot swapping, and resource scheduling.

Production reliability: Implement observability (DCGM, Prometheus, Grafana), tenant isolation, overload protection, failure recovery, and safe rollout and rollback.

Cross-environment deployment: Build portable runtime integrations across NVIDIA and AMD GPU pla4orms, cloud and on-premises environments, and single- or multi-GPU configurations.

Technical Expertise

Model inference internals: Autoregressive generation, prefill vs. decoding, attention architectures, compute- vs. memory-bound execution, numerical precision.

Speculative decoding: Draft-and-verify execution, acceptance/rejection sampling, adaptive speculation, and throughput tradeoffs.

GPU kernels: Flash attention-style kernels, tiling, operator fusion, memory access patterns.

CUDA, HIP, Triton, or CUTLASS proficiency.

KV and prefix caching: Paged KV storage, fragmentation, cache quantization, offloading, cross request prefix reuse, tenant isolation.

Context-window management: Token budgeting, context selection, prompt compression, conversation compaction, sliding window and sparse attention.

Serving and scaling: Continuous batching, chunked prefill, admission control, quantization, multi-GPU parallelism, prefill/decode disaggregation.

Qualifications Required:

• 5+ years engineering, including 3+ years hands-on with GPU systems, distributed inference, or ML serving
• Production experience with vLLM, SGLang, or TensorRT-LLM
• Strong Python and C++ (C++17/20); ability to read and debug CUDA C++ or HIP kernels
• Hands-on GPU profiling and performance optimization

Preferred:

• PyTorch internals — torch. Compile, custom operators, MoE or disaggregated prefill/decode
• LoRA/SLM adapter serving and zero-downtime hot swapping
• NVIDIA and AMD Instinct hardware in production
• Telemetry with DCGM, AMD SMI, Prometheus, and Grafana