Your software works. Their floor routed around it.
For technology vendors
In the sales demo, the data was clean and the workflow was linear. Sixty days after go-live at the customer’s site, the receiving clerk is typing dummy quantities to clear the dock, the planner has exported your screen back into Excel, and the executive sponsor is asking why the promised ROI has not shown up before renewal.
Most rollouts do not fail in the model or the core code. In a 2015 NeurIPS paper, Sculley and colleagues noted that a mature machine-learning system might end up at most 5% model code and at least 95% glue code; on tau-bench, the 2024 benchmark from Yao and colleagues, even frontier language-model agents pass fewer than 50% of multi-step retail and airline workflows when they must follow domain rules across turns. In the field, that 95% is where the customer’s dirty master data, local incentives and shift shortcuts quietly break your product.
Sources: Sculley and colleagues, Hidden Technical Debt in Machine Learning Systems, NeurIPS 2015; Yao and colleagues, tau-bench, 2024.
What I check first
- Silent corruption in the customer’s feed: before your engine runs a single calculation, I check whether one-to-many purchase-order schedule joins, un-netted in-transit inventory, or quarantined dock lots are inflating the stock numbers your ingestion pipeline reads.
- The shadow spreadsheet beside the terminal: what the buyer or supervisor calculates outside your tool, and which real physical constraint your configuration left out: minimum order multiples, dock staging limits, supplier lead-time variance, or shift cutoffs.
- The incentive collision: whether the customer’s floor KPIs punish operators for following your workflow, turning every exception into an unrecorded bypass.
- Unguarded write paths and rule loopholes: overly broad permissions, shared floor logins, unbounded live ERP writes without TTL or rollback, and untrusted external text inputs that can steer agentic workflows past your guardrails.
Where I have stood
- On the customer’s floor: thirteen years in Chile, Argentina and Australia across retail planning for more than 200 stores, a three-site logistics network of about 100 people, a USD 30M cross-border e-commerce P&L, and the warehouse-system cutover of a 50,000-SKU distribution centre where I stayed on the floor until the workarounds died and order processing time fell about 20%.
- At the keyboard building deterministic guardrails: I learned to code in C and build in Python and TypeScript with AI coding agents. I have built an inventory and replenishment engine with 10,756 passing tests, receiving and field-audit kits (1,141 and 1,187 passing tests), and a weightlifting planning application (about 1,332 unit tests and 31 PostgreSQL integration tests), staging recommendations in draft beside the operator’s own play with HMAC-SHA256 sign-off and rollback, so a bad upstream record never writes directly to production.
What I don’t claim
- I do not build your connectors or pipelines; your forward deployed engineer does. What I hand over is what the engineer can test: which records to distrust before any calculation, which writes may reach the live system (signed, time-boxed, with a way back), what the floor is paid for, and how a screen has to behave on the night shift. In the strategist and engineer postings I coded in November 2026, each of those four is missing from most of them, and none asks the role to work on the incentive. Doing all four next to the engineer is my proposal, not yet my track record.
- I have not served as a vendor’s commercial account executive, and my own diagnostic engines are field instruments rather than a competing SaaS product. What I bring to a vendor team is thirteen years on the buyer’s side of the table plus the engineering discipline to turn messy floor exceptions into a clean data contract.
- I use AI to accelerate specification drafting, boilerplate and adversarial code review; floor discovery, causal diagnosis, the mathematical invariants and the sign-off stay in my hands.
- Open to full-time or embedded Deployment Strategist roles on site with relocation or remote (if you are hiring). On fixed-scope engagements invoiced from Santiago, Chile, hands-on engineering, data-contract work and forensic analysis are performed remotely, while short on-site visits focus on floor observation, blind counts and stakeholder alignment under standard business-visitor rules.
How we would start
- Discover on the floor. Walk the customer’s receiving dock, aisles and planning desk alongside your team; reconcile a blind sample of physical stock against the ingested feed. Out: floor-to-schema gap log and baseline discrepancy count.
- Tune. Translate the customer’s real physical constraints and exception rules into a tested data contract: field tolerances, join invariants, unit-of-measure rules and fallback behavior. Out: executable data contract and edge-case test fixtures for your engineering team.
- Orchestrate. Run your system in a safe draft stage where the operator’s play, your engine’s recommendation and the status quo sit side by side before write-back goes live. Out: signed-off cutover checklist, rollback triggers and shift playbook.
- Value case. Measure the operational baseline before cutover and re-measure 30 and 90 days after go-live in metrics the customer’s CFO and COO both accept. Out: before-and-after operational ledger for renewal and expansion.
- Let go. Train the customer’s floor leads on the exception loop and hand your deployment engineers a repeatable onboarding runbook. Out: customer sign-off and the one-page onboarding plan (PDF).