A system's own status report is a claim, not evidence.

Senior product and operations leader — 15+ years of P&L turnarounds, PMO leadership and
strategic portfolios in retail energy — now building the tooling that checks
whether AI systems are telling the truth about themselves.

averyresume.com  ·  LinkedIn  ·  Book 30 minutes  ·  andrew@averyemail.com

Each of these began the same way. A tool reported a result, and the result was not true.

The first two were other people's software. The third was mine, which is the harder version of the same discipline.

Every LLM cost calculator on the web is a snapshot: someone hard-codes a dozen prices and within months the premise is wrong. PromptSpend is built the other way round — the pipeline is the product. Every morning a GitHub Action re-checks 69 models across 12 providers, merges sources under an explicit trust order, and a run that loses a source or would shrink the catalogue publishes nothing and fails loudly.

The differentiator is provenance, not breadth. Competitors advertise more models. Here every price carries the vendor page it was read from and the date it was read, and prices two sources disagree about are marked rather than quietly resolved. 43 of the 69 have been read against a vendor's own page; the other 26 stay labelled as feed-sourced, because using an aggregator to close that gap would launder a third party's number into a first-party claim.

Answers in four places — the site, a keyless CORS-open API with an OpenAPI 3.1 description, an MCP server on npm so a coding agent gets the same figures with the same paperwork, and a VS Code extension that annotates a model id inline with its rate and the date it was last confirmed. 737 automated tests across six suites, 112 of them in a real browser across four viewports, axe auditing at WCAG 2.1 A/AA. No accounts, no analytics, no cookies.

_{2 minutes 8 — press play, and hit 🔊 to unmute (GitHub starts videos silent).}

Converts the PDFs and Office files you mention to markdown and hands Claude Code a pointer rather than the contents, so referencing a large document costs almost nothing until its contents actually matter.

Because the underlying converter exits 0 on an empty extraction, every conversion is graded on the text it actually recovered — PDFs by characters per page against a threshold chosen from measured documents. Image-only PDFs measure 0–40 characters per page; real text PDFs measure 650–950; the default threshold of 100 sits in the gap between them. A failed extraction writes no file and tells Claude to read the original with vision instead. 41 tests, twelve formats, CI across Windows, macOS and Linux on Python 3.10 and 3.12. No OCR, no server, no telemetry.

One command moves a running Claude Code session into OpenAI Codex with the conversation, skills, operating instructions and memory layer intact — and opens the Codex desktop app directly on the transferred thread. A desktop launcher does the same without spending a Claude token, which matters when you are rate-limited.

The Windows false-failure diagnosis above was submitted upstream as codex-plugin-cc#551: a 19-line fix, unit-tested, verified against a real Windows 11 ledger. It was closed without review after the repository went quiet; issue #513 is still open and the bug is still there. The fix ships here instead — this kit detects the imported thread directly in Codex's state database, because it trusts the evidence rather than the success message.

Fifteen years of running operations where being wrong is expensive, which is where the habit comes from.

• Returned a $25M annual-loss business unit to profitability in 12 months at NRG Energy.
• Took $91.6M out of the cost base in three years against a four-year plan at Direct Energy — beating the target by $12.8M, 14% over.
• Stood up a PMO from scratch across multi-segment GTM partnerships; managed five team leads across ~30 staff.
• Served on the executive leadership team accountable for NRG's M&A divestiture, which unified the operating model and consolidated cross-functional teams.
• Redesigned customer engagement at Direct Energy: retention +11%, revenue +$100K in the first month.

Currently Senior Manager, Business Performance & Strategic Initiatives at NRG Energy (Fortune 150), running the strategic-initiative portfolio for a multi-segment retail energy and home-protection P&L, working with the C-suite on prioritization, capital allocation and performance reporting — and expanding that remit into evaluating generative-AI use cases against ROI, accuracy and risk criteria.

Post-graduate AI & Machine Learning: Business Applications at UT Austin (McCombs), in progress — four graded builds spanning multi-agent automation, RAG document analysis, predictive-maintenance networks and clinical decision support. Executive education in AI strategy at UC Berkeley (Haas). 19 AI certifications alongside CSM, CAL-E and CPP.

Ships code daily with Claude Code and Cursor, with evidence-based judgment about where LLM-assisted development helps and where it does not — evidenced by the three repositories above rather than asserted.

Open to senior product, operations and AI leadership roles and contract engagements.
Industry-open.

Book 30 minutes  ·  andrew@averyemail.com  ·  Résumé  ·  LinkedIn