ServiceNow & SAP Labs: ML / AI Engineer

def max_profit(costs, profits, budget):          # 0/1 knapsack
    dp = [0] * (budget + 1)
    for c, p in zip(costs, profits):
        for b in range(budget, c - 1, -1):
            dp[b] = max(dp[b], dp[b - c] + p)
    return dp[budget]
print(max_profit([5, 4, 6, 3], [10, 40, 30, 50], 10))   # 90

• This is 0/1 knapsack: O(n·B) time, O(B) space. If the problem is instead "buy and sell one stock over time", it is a single pass tracking the minimum price.
• Clarify assumptions first (integer costs? one share each?). The candidate's frustration came from mismatched assumptions.

• Inputs: the ticket text, the user's role, CMDB (asset) data, past resolved tickets, the knowledge base.
• Flow: 1. Classify the category and priority. 2. Retrieve similar resolved tickets and KB articles (hybrid search). 3. Draft a resolution or a suggested action. 4. For safe automations (password reset, access request), call workflow tools with approval policies.
• Guardrails: role-based tool permissions, a human agent in the loop for changes, an audit trail.
• Metrics: deflection rate, mean time to resolve, reopen rate, agent acceptance of suggestions.

• Supervised: predict labelled outcomes, such as invoice approval (classification) or demand forecasting (regression).
• Unsupervised: find structure, such as vendor clustering or anomaly detection in journal entries.
• Reinforcement: learn by reward, such as inventory replenishment policies or dynamic pricing.
• Self-supervised / foundation models: LLMs for document understanding, a copilot over ERP data.

• Structured data: use tools/APIs (OData services) or text-to-SQL over a curated semantic layer, rather than dumping tables into prompts.
• Unstructured data: RAG over policies and documentation.
• Authorisation: enforce the user's existing business roles on every tool call; never give the model broader access than the user has.
• Accuracy: calculations go through deterministic services; answers cite the source records; confirmation before posting transactions.
• Compliance: data residency, audit logs, no customer data used for training without consent.

• Open with a 30-second summary (problem, users, result), then show the architecture diagram, then a live demo of the core flow, then the code for one interesting part (for example, the retrieval logic).
• Have a fallback (screenshots or video) if the demo fails. Know your repo structure well enough to navigate instantly.

• S/4HANA is SAP's ERP suite on the in-memory HANA database, covering finance, supply chain, procurement and manufacturing.
• For AI: it is the system of record that copilots and predictions act on (cash-flow forecasting, invoice matching, supply-risk prediction). Understanding business processes (order-to-cash, procure-to-pay) lets you build AI that fits real workflows.