Estimated duration
~4h
Deadline
Wed, Sep 30, 2026
Location
📍 India (Remote OK)
USD
Human Company builds expert-grade reinforcement-learning environments used to train and evaluate AI models on inference engineering. We are seeking an ML Systems Engineer to own inference-systems tasks end to end: choose a realistic objective, build the execution environment, implement robust ground-truth verifiers, and perform rigorous QA on fidelity, ambiguity, exploit resistance, and difficulty calibration.
This bounty is a paid, four-hour technical work sample for candidates based in India. Strong submissions may lead to a remote full-time or monthly contract role with a broader compensation range of ₹22-110 LPA, calibrated to experience and scope.
Strong candidates have operated production inference systems and owned latency, throughput, or memory outcomes. Depth in at least one of the following is required, with working knowledge across the rest: serving stack internals (continuous/in-flight batching, chunked prefill, scheduling, admission control, KV-cache management); model efficiency (quantization, speculative decoding); distributed execution (tensor, pipeline, sequence, and expert parallelism, disaggregated prefill/decode, multi-replica balancing, elastic inference); hardware and compilation (GPU/FPGA/ASIC kernels, TensorRT-LLM, TVM, roofline-guided optimization); and statistically sound profiling and benchmarking. Contributions to vLLM, SGLang, TensorRT-LLM, or comparable systems are especially relevant.
The accepted candidate will produce a compact design and verifier prototype for a realistic inference-engineering RL task. Do not share proprietary code or confidential employer information.
Deliver a compact, expert-grade inference-engineering RL task package that can be reviewed in approximately four hours of work. It must include: (1) a fair, unambiguous task prompt grounded in a realistic serving-stack scenario; (2) a concise description of the reproducible execution environment; (3) a runnable verifier prototype or precise pseudocode plus representative tests defining success, failure, constraints, and important unspecified behavior; (4) a benchmark or profiling plan that controls for noise; (5) QA notes covering ambiguity, loopholes, verifier robustness, difficulty calibration, and known limitations; and (6) clear reproduction or review instructions. Do not include proprietary code or confidential employer information.
EVIDENCE DETAILS
Provide a written summary and a link to a public or owner-accessible private repository, gist, notebook, or document containing the task design, verifier prototype or precise pseudocode, representative tests, benchmark plan, QA analysis, and reproduction/review instructions.
Based in India and able to work remotely
Production ownership of model-serving latency, throughput, or memory
Deep expertise in at least one inference-systems domain described in the listing
Able to design reproducible environments and robust, exploit-resistant verifiers
Strong written technical communication
USD
Create a free account to apply for this bounty