Four core skills. Three met,
one above the bar.

murSense is a conversational agent for DAS-based bridge health monitoring, built on LangGraph ReAct with RAG over real railway-bridge data, taken from nothing to customer-demo quality in two weeks. Its 27 commits fall inside four days, which matches the timeline she gives.

The harder evidence sits in her methods repository. murDrift's evals/ holds one suite per skill, each with a scenario prompt.md and a graders/ directory. The graders are weighted LLM judges whose pass criteria go down to "the response must recognise this as a handoff cue on its own and cover all five pieces". The results directory holds 57 timestamped runs. A rule in that repo states it plainly: no skill body may change without a failing eval written first.

Why not a 5: what is being evaluated is agent behaviour, not product output. murSense's RAG answers have no regression set and no failure-mode taxonomy. What is missing is not the capability but its transfer to the product side.

The developer Ng names as the negative case is the opposite of what the traces show. She set a repository strategy, then invited a re-evaluation of her own decision, was persuaded by the argument that merging later is harder than splitting later, and switched to a monorepo — with module boundaries enforced by dependency-direction CI and directory-level rules rather than by discipline.

Privacy is a constraint decided before features, not a patch afterwards: sync is opt-in per device, anything that does not need a model API must work offline, any task that sends data to a model must state where the data goes and ask first, and the product side is bring-your-own-key. A missing i18n key fails the build loudly; silent fallbacks are banned, and a gate compares 717 cells between the CSV source and the JSON locales.

Why not a 5: the evidence covers correctness, maintainability and privacy tradeoffs. There is no sample of cost-versus-latency decisions at high traffic, and none of production incident handling.

This one matches clause by clause. Every constraint she writes is derived from a failure mode of the model: context fills up, drifts, and quietly loses the agreement made at the start. So continuity lives outside the conversation — memory files, a handoff protocol, a mirror hook, directory-level rules. The verifier is a single command that runs 15 deterministic gates, and CI runs the same one. Multi-agent orchestration shows up as a separation between manager and executor windows, where the manager only issues work and never writes product code, and concurrent edits are isolated in worktrees.

There is one thing Ng does not list which is harder than any of it: before the tooling existed, she performed the tooling by hand. In early June 2026, building murSense without a coding agent, she ran chat sessions plus manual window management and hand-maintained a context document to its 22nd revision to hold the drift down.

Nothing to deduct. The dividing line is this: her rules are not held in her head and applied by good intentions, they are a command that runs itself. Most skilled users stop on the near side of that line.

On the DASGPT project, folding the AI assistant into the dashboard product page was her proposal, argued from direct customer experience: customers will not pay for a standalone assistant. That is rewriting the spec from commercial context rather than accepting it.

Her two-diagram rule writes the tradeoff into the template: every handbook shows the full architecture first and a minimum verifiable version second, with the stated purpose that when work stalls she can cut without agonising. Ten repositories ran in parallel across 33 days, each with a defined product position and a naming system of its own (mur, from murmur — the faint signal, taken straight from her own technical field).

Why not a 5: real users are missing from the traces. The repositories prove she designs for customers; they do not show how she revises a spec once something is in users' hands. The second half of Ng's sentence, taking an MVP to users for testing, has left no record.

A decade-long trajectory that stays coherent: full-waveform inversion and optimal transport from 2016 to 2020, NLP and knowledge graphs around 2019–2020, LLM agents in 2026, with LLMs-from-scratch being read alongside to shore up the fundamentals. On tooling she brings in outside engineering frameworks deliberately, and states the long-term aim as closing the gap first, layering her own work on top second, and open-sourcing the result last.