Short note on Atlas, a 2021 private service I built for bursting quantum-optimisation jobs from a small control plane to AWS Braket (D-Wave and gate-model backends). Jun–Nov 2021, ~59 commits. Closed source.

What it did

A minimal API + auth layer + Docker-on-EC2 executor:

• Auth and credit: cmdAddUser, cmdAddCredit, cmdGetToken. Token-based API; credits metered against run cost.
• Burst: build.burst.portfolio.sh, burst-monitor, burstbuild.sh / burstbuild.expect. A client submits a job; the controller packages it into a Docker image configured via ATLAS_SECRET, ships it to EC2, and streams logs back.
• Quantum backends: BraketDwave.py, BraketQAOA.py. D-Wave annealer path for QUBO problems; QAOA path for gate-model simulators.
• QUBO / Ising plumbing: ContinuousToBinary.py, cov_to_qubo.py, VQE_to_ising.py, multi-bit resolution encoding so a continuous-valued portfolio weight could land in a binary-variable QUBO cleanly.
• RL demo: an early reinforcement-learning portfolio demo (Jul 2021, “better RL demo”) that depended on LaplaceKorea/Portfolio-Optimization-and-Goal-Based-Investment-with-Reinforcement-Learning.
• Ising experiments: ising commit in Jul 2021 — classical Ising ground-state checks against the quantum/annealer results.

Why burstable

At the time the question was whether a quantum backend was worth the latency for a specific class of portfolio problems. The only honest way to answer that was to run the same QUBO through an annealer, through QAOA on a simulator, and through classical solvers, and compare. Atlas was the plumbing that made “run the same problem on all three” one command.

The answer, for the problem sizes I could run, was “not yet, for this family of problems.” That’s the kind of answer you want to be able to produce, even if it’s negative.

What I’d do differently

• Don’t build an auth system. A pre-shared key on top of SSH port-forwarding would have been fine for the audience.
• The multi-bit-resolution conversion is the interesting piece. Promoting that into a standalone library — “continuous-to-binary for QUBO” with a clear error bound per bit — would have had more reuse value than the whole burst apparatus.

Closed source. The BraketDwave.py / BraketQAOA.py shapes later informed the QuetzalcoatlProto ansatz-benchmark harness.