Founding ML Infrastructure Engineer
Kalpa Labs (company site, opens in a new tab)3+ yrsBengaluruOn-siteDirect career pageEst. 18–30 LPA
Posted 1d ago
Reach people at Kalpa Labs
People at this company you can reach
Log in to see who works here
Create a free account or log in to see the people at Kalpa Labs you could write to:
- Recruiters & HR
- People in this function
- Work addresses
Your match
Against your primary resume
Log in to see your match score
Create a free account or log in to unlock your CV match score across:
- Role fit
- Experience
- Industry relevance
Protocol Intelligence
How competitive this role is
Competition
Background fit
Strictness
Log in to see why each signal reads the way it does.
Job description
Structured overview of role & requirements
About This Role
- Own and optimize the distributed training stack including FSDP/HSDP and Megatron-style tensor parallelism in PyTorch.
- Improve realtime inference performance focusing on latency and throughput using vLLM-based engine, quantization, model compilation, and high concurrency load testing.
- Build and maintain high-throughput data pipelines processing millions of hours of audio and manage production APIs including streaming WebSocket APIs, deploy tooling, logging, and monitoring.
Minimum Requirements
- 3+ years experience building ML systems or high-performance backend infrastructure.
- Hands-on expertise in at least one: distributed training (FSDP/TP/PP), LLM serving (vLLM/SGLang/TensorRT-LLM), high-volume data processing, or low-latency streaming systems.
- Work Experience Required: 3+ years in relevant fields.
- Location: Bengaluru (implied onsite requirement).
Ideal Candidate Profile
- Experienced building scalable ML infrastructure components end-to-end, from training to serving to production deployment.
- Comfortable working independently in ambiguous, fast-moving environments while spec'ing, building, debugging, and shipping complex ML system components.
- Prefers candidates with domain exposure to speech/audio ML infrastructure but not mandatory.