| Your Boss | Matt Douglas — 4 direct reports. Bishop is one of four. |
| Naaga | Your peer. Principal SW Eng. Evaluating fit, not gatekeeping. |
| Ladder | Principal → Sr. Principal → Distinguished Engineer |
| Title | Principal Software Engineer |
| Tenure | 7y 7mo at HD |
| Role Type | Peer Validator |
| Top Skills | Informix 4GL, Hadoop, Spring MVC |
| Background | TCS consultant at HD SSC → Senior → Staff → Principal. Knows the legacy stack cold. |
| Her lens | Can Bishop design real systems? Will he fit the HD engineering culture? Is he operational-minded or academic? |
"I've spent 40 years building systems where failure has immediate patient or customer consequences. Nuclear pharmacy PET isotope delivery runs on 2-hour windows — miss it, and the imaging suite has nothing to inject. Carrier cut times are the same problem type. The physics is different. The constraint structure is identical."
"AI today is 1–2 people." This is greenfield. Naaga will probe whether Bishop can define how AI gets done — not just implement someone else's pattern.
In scope: Web order → customer home delivery OR store pickup. Two pick systems: robotic (small items → tote → conveyor → station) and human pull (large items, appliances, pallets → truck).
Out of scope: SKU to store replenishment, flatbed/lumber to job sites.
Hard cuts: FedEx 10am + 2pm. Everything optimizes around these two windows.
"I'd decompose it into four layers, each with its own latency budget."
Layer 1 — Eligibility gate (<100ms): Rule-based constraint check. Which orders are eligible right now? In stock, cut time not yet breached, station capacity available. Fast filter — no ML here.
Layer 2 — Scoring model (200–500ms): Supervised ML on tabular features — time to cut, items per tote, zone congestion index, station load, SKU co-location. Output: urgency-adjusted throughput score. Not an LLM.
Layer 3 — Batch formation (500ms–2s): Select top-N from ranked list that minimizes robot congestion. Greedy or beam search. N = available station slots.
Layer 4 — Station assignment (<50ms): Greedy assignment to least-loaded eligible station. Simple, auditable, reversible.
Layer 5 — Feedback loop (async): Actual pick time vs predicted → retrain scoring model. Cut-time compliance rate → rollback trigger if below threshold.
"The key design principle: each layer has a different latency requirement and a different update cadence. Don't collapse them into one system."
"Carton streaming is a structured decision problem — structured inputs (order IDs, item locations, cut times, robot positions), structured outputs (binary pick/no-pick, station assignment). That's supervised ML, not LLMs."
"LLMs belong where the input is unstructured. An associate verbally reports a jam or an inventory exception — that's a natural language → structured action problem. LLM handles the interface, ML handles the decision."
"The question I always ask: what's the latency budget? Hot path decisions every few seconds eliminates LLMs entirely — inference latency and cost alone rule them out. LLMs sit in the exception handling layer, not the execution layer."
"And interpretability matters here. If the system makes a bad batch decision and a cut time is missed, someone needs to explain why. A tabular ML model with logged features is auditable. An LLM is not."
"Shadow mode first. Run the AI model in parallel with the rules-based system — don't act on AI decisions yet. Measure: would the AI have decided differently? When it did, would it have been better?"
"Define 'better' before you run the experiment: throughput rate, cut-time compliance %, robot travel distance, station idle time. No undefined success criteria."
"After shadow mode validates direction, controlled A/B at low exposure — 5 to 10% of decisions. Same metrics. Set rollback triggers: if cut-time compliance drops below threshold, revert automatically."
"The rules-based system is not decommissioned — it's the fallback hot path. You don't delete the thing that's working until you've proven the replacement is better across all dimensions, not just the headline metric."
"I've run 27 controlled iterations of this type of experimental cycle at NewsRx — systematic, tracked, measurable. Same discipline applies here."