Design PII-safe call summarisation for an Indian contact centre

    Medium35 minSystem design · AI rubric, per criterionCFDE Pro
    • PII constraint
    • cost/latency budget
    • structured output

    The brief

    Suraksha Life (fictional), a life insurer, runs a contact centre in Noida and Chennai. Agents currently spend about 90 seconds after each call typing notes into the CRM, and the notes are often incomplete.

    The client wants every call summarised automatically into structured CRM fields: reason for call, policy number, actions promised, follow-up date, complaint flag, and a 3-line summary.

    Constraints:

    • 85,000 calls a day, average 5.5 minutes, peak 9,000 calls an hour. Languages: Hindi, English, Tamil, Marathi, frequently mixed within one sentence.
    • Summary must appear in the CRM within 2 minutes of call end.
    • Customers sometimes read out card numbers and CVVs during premium payments. These must never be stored in transcripts, logs or summaries (PCI DSS). Aadhaar and PAN numbers also come up.
    • Personal data must stay in India under the client's reading of the DPDP Act and their regulator's outsourcing guidelines. Call recordings already sit in an India-region object store.
    • All-in budget: ₹2.50 per call, including speech-to-text.
    • The complaint flag feeds a regulatory complaints register; missing a complaint is far worse than a false flag.
    • The CRM accepts updates through a REST API with a fixed JSON schema.

    Design the system.

    How it's graded

    • Sensible pipeline from audio to CRM2 pts
    • Layered PII/PCI handling2 pts
    • Cost and latency maths that fits ₹2.50/call and 2 minutes2 pts
    • Treats complaint detection as a high-recall problem2 pts
    • Evals and rollout2 pts

    Your answer

    The end-to-end flow: where data enters, where models run, what is stored where. A text diagram is fine.

    0/250 words

    Models, retrieval, tools, guardrails. Say why you picked each and what you rejected.

    0/250 words

    Take the stated constraints one by one, with numbers where they apply.

    0/220 words

    How you know it works before and after launch, and how you roll it out safely.

    0/200 words

    Graded by AI against the ADaSci rubric. Your answers stay private; only your score shows on your profile.