A Solo Engineer Placed 12th Out of 183 With a 232x Kernel Speedup Using Codex. The Playbook: Sell 'Autonomous Optimization Harness' Builds, Not Just the Code.

A blog post walking through a GPU kernel optimization contest just made the Hacker News front page — not because of the math, but because of the process.

The author entered GPU Mode's auto-research contest (run with Core Automation), where the task was to implement batched square compact-Householder QR factorization. The result: "I placed 12th out of 183 participants, ending up with a 232x speedup over the baseline solution." Over 14 days he made "over 1500 submissions," and the tracked lineage on his best line shows the arc from 108,803 µs down to a final 1,805 µs.

The business idea

Read past the kernel math and there's a repeatable service underneath: building the harness that lets a coding agent optimize autonomously, instead of a human hand-tuning code line by line.

The author's setup was simple and documented in detail — a problem_statement.md to pin the target, an AGENTS.md with "basic details... on how to submit," and a log.md for bookkeeping so "future agent sessions could read the logs and quickly check if an idea had been tried or not." He gave the agent quantitative goals like "beat our active best's n = 512 timings," then let it run loops autonomously, checking in without interrupting the loop via commands like /btw.

The harder part — and the real service opportunity — showed up once easy gains ran out. In his words, "a major challenge I started facing in the 3000 -> 1800 µs range was the model getting stuck in local maxima." His fix: stop hill-climbing on a single best candidate and instead "maintain a beam of 3-5 candidates," use "a stronger advisor model that produces more varied ideas" for headless second opinions, and deploy sub-agents to hunt for bugs and micro-optimizations in parallel.

That's not a script you download once. It's a methodology — and most engineering teams sitting on a slow kernel, a slow query, or an expensive inference path have neither the time nor the harness discipline to build it themselves.

What to sell

Package this as a fixed-scope "autonomous optimization harness" build: set up the problem_statement/AGENTS/log structure for a client's specific bottleneck, wire in an advisor-model pattern for when the primary agent stalls, and hand over a working loop plus the documented submission history as proof of the ratchet. Price it against the client's own before/after number, exactly the way this post's pitch is really just one line — 12th of 183, 232x — that anyone can verify.

Turn it into a retainer by re-tuning the harness whenever the client's workload shifts or the underlying model changes. The author's own conclusion points at where this is headed: "the strong advisor strategy... is going to be a standard strategy in auto-research flows."

Who buys this

ML infra teams, backend teams with a slow hot path, and any engineering org already paying for coding-agent subscriptions but using them as autocomplete rather than as an autonomous optimization loop with real submission discipline.

Bottom line

The 232x number is the hook. The actual product is the harness that produced it — and that harness is exactly the kind of thing most teams will pay someone else to build correctly the first time.

Source: https://sankalp.bearblog.dev/autoresearch/