Startup seeking a senior freelance JAX/Scientific Computing Engineer to implement a locked mathematical specification in production JAX. Math already designed by an applied mathematician. Remote, 20-30 hrs/week, long‑term contract. NDA required before detailed spec access.
Technical surface (you don’t need all of it on day one, but you’ll touch most of it):
-
High‑dimensional square‑root unscented Kalman filtering (Joseph‑form updates, adaptive noise)
-
SDE integration (Euler–Maruyama, Milstein, exact‑transition schemes)
-
Hierarchical Bayesian inference via NumPyro NUTS (running pre‑specified models, not designing them)
-
Amortized neural posterior estimation via conditional normalizing flows (BayesFlow)
-
Multi‑state hazard modelling, doubly‑robust estimators, constrained contextual bandits
-
JAX → ONNX export pipeline (jax2tf → tf2onnx, golden‑set parity testing)
Stack:JAX, Equinox, Diffrax, NumPyro, BayesFlow, Triton Inference Server, vLLM (co‑tenant on GPU node), Dagster, FastAPI, ONNX Runtime (on edge devices). Single H100‑class GPU, scaling planned.
What we need :
-
Production JAX experience (vmap, lax.scan, JIT compilation, FP8/BF16 memory budgeting)
-
Familiarity with Equinox or Flax, Diffrax (or another ODE/SDE library), NumPyro
-
At least one model‑export pipeline (jax2tf + tf2onnx, or equivalent)
-
Ability to read a math spec and turn it into efficient, tested JAX code without debating the upstream design choices
What this is not:
-
Not a math‑design role (priors, estimators, and model structure are locked)
-
Not a PyTorch job repackaged — we need genuine JAX‑internals fluency
Engagement:Hourly contract, 20-30 hrs/week, 4‑month initial term with strong intent to continue. Rate $80–100/hr depending on background. Fully remote, async‑first. Weekly 30‑min architecture sync.
Next step:If interested, please reply here or email me at alicharanek@gmail.com with a few lines about your JAX experience. We’ll schedule a 30‑min call, and if it’s a mutual fit, we’ll share the full spec under NDA.
Startup seeking a senior freelance JAX/Scientific Computing Engineer to implement a locked mathematical specification in production JAX. Math already designed by an applied mathematician. Remote, 20-30 hrs/week, long‑term contract. NDA required before detailed spec access.
Technical surface (you don’t need all of it on day one, but you’ll touch most of it):
High‑dimensional square‑root unscented Kalman filtering (Joseph‑form updates, adaptive noise)
SDE integration (Euler–Maruyama, Milstein, exact‑transition schemes)
Hierarchical Bayesian inference via NumPyro NUTS (running pre‑specified models, not designing them)
Amortized neural posterior estimation via conditional normalizing flows (BayesFlow)
Multi‑state hazard modelling, doubly‑robust estimators, constrained contextual bandits
JAX → ONNX export pipeline (jax2tf → tf2onnx, golden‑set parity testing)
Stack:JAX, Equinox, Diffrax, NumPyro, BayesFlow, Triton Inference Server, vLLM (co‑tenant on GPU node), Dagster, FastAPI, ONNX Runtime (on edge devices). Single H100‑class GPU, scaling planned.
What we need :
Production JAX experience (
vmap,lax.scan, JIT compilation, FP8/BF16 memory budgeting)Familiarity with Equinox or Flax, Diffrax (or another ODE/SDE library), NumPyro
At least one model‑export pipeline (jax2tf + tf2onnx, or equivalent)
Ability to read a math spec and turn it into efficient, tested JAX code without debating the upstream design choices
What this is not:
Not a math‑design role (priors, estimators, and model structure are locked)
Not a PyTorch job repackaged — we need genuine JAX‑internals fluency
Engagement:Hourly contract, 20-30 hrs/week, 4‑month initial term with strong intent to continue. Rate $80–100/hr depending on background. Fully remote, async‑first. Weekly 30‑min architecture sync.
Next step:If interested, please reply here or email me at alicharanek@gmail.com with a few lines about your JAX experience. We’ll schedule a 30‑min call, and if it’s a mutual fit, we’ll share the full spec under NDA.