Data Scientist - Model Optimization at quadric, Inc

Burlingame, California, United States -

Full Time

Start Date

Immediate

Expiry Date

01 Jul, 26

Salary

0.0

Posted On

02 Apr, 26

Experience

5 year(s) or above

Remote Job

Yes

Telecommute

Yes

Sponsor Visa

Skills

Model Optimization, Quantization Algorithms, Python, PyTorch, TensorFlow, NumPy, Pandas, SciPy, Data Visualization, Fixed-Point Arithmetic, Numerical Analysis, Statistical Calibration, CNNs, Transformers, DNN Architectures, Custom Hardware Accelerators

Industry

Semiconductor Manufacturing

Description

Quadric has created an innovative general purpose neural processing unit (GPNPU) architecture. Quadric's co-optimized software and hardware is targeted to run neural network (NN) inference workloads in a wide variety of edge and endpoint devices, ranging from battery operated smart-sensor systems to high-performance automotive or autonomous vehicle systems. Unlike other NPUs or neural network accelerators in the industry today that can only accelerate a portion of a machine learning graph, the Quadric GPNPU executes both NN graph code and conventional C++ DSP and control code. Role: You will be joining the data science team focused on model optimization for Quadric's custom GPNPU architecture. You will research, prototype, and implement novel quantization algorithms tailored to our hardware constraints. Beyond applying existing techniques, you'll develop custom low-precision methods that maximize performance on the Chimera GPNPU. Your work will directly shape the quantization capabilities in the Chimera SDK and influence future hardware features. This California Bay Area based engineering role is intended to be primarily in-office at our Burlingame location, with the ability to commute regularly. We believe strong technical collaboration, rapid iteration, and shared problem-solving are well supported by working together in person. The team and company also gather periodically for onsite meetings and offsite events to connect, collaborate, and align on priorities. Responsibilities: Design statistically rigorous experiments to compare PTQ, QAT, and mixed-precision schemes on vision, language, and multimodal models. Implement custom quantization algorithms from scratch, adapting existing techniques or developing novel approaches to match Chimera GPNPU's unique architectural features and numerical formats. Build calibration datasets; develop Python notebooks/dashboards to track accuracy, latency, power, and memory trade-offs. Perform layer-level error analysis to guide numerical-format choices. Partner with compiler team to convert your findings into turnkey SDK flows and reference configs. Publish internal white papers, external benchmarks, and present results to customers and at industry events. Monitor academic literature in compression and efficient inference; translate promising ideas into reproducible prototypes. M.S./Ph.D. in CS, EE, Applied Math, or similar, with 5+ years in ML model optimization or data-science-driven research. Deep grasp of fixed-point arithmetic, quantization theory, numerical analysis, and statistical calibration. Strong ability to implement quantization algorithms from first principles, not just use existing frameworks. Fluent in Python, PyTorch or TensorFlow, NumPy/Pandas/SciPy, and data-viz tools (Matplotlib/Plotly). Experience implementing custom quantizers and understanding their interaction with hardware constraints (bit-width, format, operations). Hands-on with at least one quantization toolkit (PyTorch FX/PTQ/QAT, TF-Lite, ONNX-Runtime, TVM, MLIR Quant) and ability to extend them. Working knowledge of CNNs, Transformers, and DNN architectures. Bonus: Experience with custom hardware accelerators, DSPs, or neural processing units. At Quadric, we value Integrity, Humility, and Happiness. What we expect from one another is simple and clear: Initiative, Collaboration, and Completion. We are a collaborative team focused on building something extraordinary in the edge computing space. Competitive salary and meaningful equity Medical, dental, and vision plan options starting on day one 401(k) retirement plan Flexible paid time off (unlimited, non-accrual) to support work-life balance When working in-office, enjoy company-provided lunches and a stocked kitchen Convenient office location within walking distance of the Caltrain station Support for commuting, including monthly parking or Caltrain passes Downtown Burlingame office location, close to shops, cafes, and local amenities A politics-free, highly collaborative environment where talented people can do their best work and make an immediate impact The opportunity to build long-term career relationships in a company that values strong personal connections alongside professional excellence Founded in 2016 and based in downtown Burlingame, California, Quadric is building the world’s first supercomputer designed for the real-time needs of edge devices. Quadric aims to empower developers in every industry with superpowers to create tomorrow’s technology, today. The company was co-founded by technologists from MIT and Carnegie Mellon, who were previously the technical co-founders of the Bitcoin computing company 21. Quadric is proud to be an equal opportunity employer. We are committed to creating an inclusive environment where people from all backgrounds can do their best work. We consider all qualified applicants without regard to race, color, religion, sex, gender identity or expression, sexual orientation, national origin, age, disability, veteran status, or any other protected characteristic under applicable law. If this role resonates with you, we encourage you to apply even if your experience does not perfectly match every qualification. We value potential, curiosity, and a willingness to learn just as much as direct experience. Skills and growth come in many forms, and we would love to hear your story.

Responsibilities

Design and implement novel quantization algorithms tailored to Quadric's GPNPU architecture. Conduct experiments to compare various quantization schemes and develop calibration datasets to track performance metrics.