A14Advanced capstone 40 to 50 hours

An observable back-office agent

A finance back-office agent for a Zambian SME that reconciles mobile-money statements against invoices, prepares Smart Invoice entries for review and chases overdue customers on WhatsApp, with every money-affecting action approved by a person and every step traceable.

The problem

A Lusaka distributor receives hundreds of MTN and Airtel mobile-money payments a month, often with a phone number and a half-typed reference instead of an invoice number. Someone spends days matching payments to invoices, preparing Smart Invoice entries and chasing late payers. It is exactly the work an agent can take on, and exactly the work where a confident mistake moves money.

You build the agent the way the source course defines agent systems engineering: a typed tool registry exposed as an MCP server, an explicit state machine, memory with rules for what persists, orchestration you can reason about, a failure and prompt-injection suite (including instructions hidden in payment references and WhatsApp replies), evaluation on trajectories rather than final answers, and runtime traces showing every stop, retry and escalation.

Architecture

statement CSV (MTN, Airtel)   open invoices   customer ledger
        \                |                 /
         +------ MCP server: typed tool registry ------+
         | parse_statement  match_payment  get_invoice |
         | draft_smart_invoice  send_whatsapp_template |
         | request_approval  (writes need approval)    |
         +---------------------------------------------+
                            |
              state graph: ingest -> match -> review -> act
                 |            |          |          |
              memory:     confidence   human     audit log
              customer    threshold    approval  + traces
              aliases                   queue
                            |
         eval: trajectory suite + injection suite + false-completion checks

What you hand in

  • A typed tool registry served over MCP, with read tools and approval-gated write tools
  • An explicit state graph with retry, stop and escalation paths
  • Memory for customer aliases and phone numbers with expiry rules
  • An injection and failure suite: instructions in payment references and WhatsApp replies, duplicate payments, partial payments
  • A trajectory evaluation on 100 labelled statement lines, and traces for a full month’s run

How it is marked

20

Trajectory evaluation

Match accuracy and correct escalations on 100 labelled lines

20

Safety and injection defence

Pass rate on the injection and failure suite

20

Runtime evidence

Traces show every stop, retry, escalation and approval

15

Tool contracts

Typed schemas, MCP registry, read and write separation

15

State and memory

Explicit transitions; memory with documented expiry

10

Human approval and audit

No money-affecting action without a recorded approval

Weights out of 100

Get it reviewed

Paste your capstone report. Jev scores the evidence against each rubric line; the final mark comes from your mentor’s code review.

Reviews open with an access code.

Adapted from open-source work: AI Engineering from Scratch by Rohit Ghumare (MIT, lesson text, code and quizzes); Awesome LLM Apps by Shubham Saboo (Apache-2.0, starter code); AI Engineering Interview Questions by Outcome School (Apache-2.0, interview questions). Capstones, milestones, data packs and Zambian context by Zambrite.

Chat on WhatsApp