Inference Performance Engineer - MaaS at Bitdeer Technologies Group in Singapore; Penang

Job description

About Bitdeer: Bitdeer is a world-leading technology company for Bitcoin mining and AI cloud. Bitdeer is committed to providing comprehensive Bitcoin mining solutions for its customers. Apart from designing industry-leading ASIC chips and manufacturing mining rigs, the Group handles complex processes involved in computing across the value chain. This includes equipment procurement, transport logistics, datacenter design and construction, equipment management, and network and facility operations. Bitdeer also offers advanced cloud capabilities to customers with a high demand for artificial intelligence. Headquartered in Singapore, Bitdeer operates globally with a diversified 3 GW energy portfolio, and deploys Bitcoin mining and HPC datacenters in the United States, Bhutan, Norway, Canada, Malaysia, and Ethiopia. About the role We are seeking an Inference Performance Engineer to make Bitdeer AI's MaaS platform's benchmarks, capacity plans, and cost projections trustworthy proxies for what customers actually run in production. Today, load tests are built on prompt and output length assumptions and concurrency patterns that were reasonable a quarter ago but no longer match the traffic mix flowing through chat, completion, and embedding endpoints — the result is GPU fleets sized against a workload that doesn't exist, cost-per-token figures that surprise finance at quarter close, and latency percentile dashboards that different teams calculate differently and therefore don't trust. This role exists to close that gap: characterizing real production traffic shapes, keeping benchmarking harnesses honest against that reality, and turning the results into capacity and unit-economics numbers the business can plan around with confidence. The person who thrives here is a rigorous, data-first systems thinker who is equally comfortable pulling apart a token-length histogram, arguing with a serving-stack engineer about tail latency, and explaining a GPU-hour cost model to finance — someone who treats every metric definition as a contract other teams will build on, not a throwaway number. What you will be responsible for: Traffic and Workload Characterization : Continuously profile real production request streams across chat, completions, embeddings, and fine-tuning endpoints to build and maintain statistical models of prompt length, output length, concurrency, and burstiness by model and customer segment. Surface how these distributions shift over time and by tenant so that benchmarking, capacity, and pricing decisions are grounded in what the platform actually serves rather than stale assumptions. Package these findings into reusable traffic profiles other teams can pull from directly. Harness-to-Production Fidelity : Own the validation loop between internal load-testing and benchmarking harnesses and live production behavior, auditing synthetic workload generators against the traffic profiles you maintain and flagging drift before it corrupts a capacity or optimization decision. Work with the inference serving team to update harness configurations, request mixes, and concurrency patterns whenever production shifts. Treat every benchmark result as suspect until it has been reconciled against real telemetry. Capacity Planning and Unit Economics : Translate workload characterization and GPU utilization data into capacity forecasts and cost-per-token or cost-per-request models that inform fleet sizing, reservation versus on-demand mix, and pricing strategy for the MaaS product. Partner with infrastructure and finance teams to reconcile modeled costs against actual cloud and hardware spend, and maintain the analytical models that answer how many GPUs are needed and what they will cost with defensible numbers rather than gut calls. Metric Definition Ownership : Define and maintain the canonical definitions for latency percentiles (time-to-first-token, inter-token latency, end-to-end), throughput, and utilization metrics used across engineering, pr