Sample
What a finished FDE project should leave behind
This is an example project written by our editorial team. It shows what deliverables look like and how a scoring report is written. It is not learner work and is not part of the assessment bank.
Example project: Lanchuan Logistics waybill exception attribution assistant
The company, people and figures are all fictional. The client is a regional logistics company that requires its data to stay on its internal network. The assessment projects have different scenarios and hidden cases. Copying this material will not help you pass the assessment.
1. Week-one interview notes
Shows you talked to the right people and surfaced the real problems in the data and the process, rather than just writing down that the client wants an AI.
- Interviewees: operations manager (the requester), dispatch team lead (the daily user), IT director (the gatekeeper for data and deployment). Three sessions of 45 minutes each.
- The operations manager asked for “automatic root causes for exception waybills.” On follow-up, the real target was customer complaint response time, currently 2 days on average.
- The dispatch lead checks three systems by hand (TMS, the GPS platform, the customer service tickets) to explain each shipment, about 6 minutes per shipment, with 150–200 exceptions a day.
- IT director: data cannot leave the internal network. The GPS platform only offers daily CSV exports and has no API. The TMS runs on Oracle, and a read-only account needs approval, which takes about a week.
- Open question: who defines “exception”? The three departments currently use different rules (over 4 hours, over 8 hours, or only when a customer complains). The operations manager agreed to decide on Thursday.
2. Data dictionary and ontology
Shows you mapped the client's business objects to their fields, and that each definition has a named owner who signed off.
| Business object | Source field | Definition / cleaning rule | Confirmed by |
|---|---|---|---|
| Waybill | TMS.WAYBILL_NO | Trim leading and trailing spaces. Voided waybills (STATUS=9) are excluded from analysis | IT director |
| Promised time | TMS.PROMISE_TIME | Originally Beijing time with no time zone. Normalized to +08:00 | IT director |
| Actual delivery | GPS export: delivery time | GPS takes precedence when it disagrees with the TMS receipt. Differences over 2 hours go to manual review | Dispatch team lead |
| Exception | Derived | Delivered more than 6 hours after the promised time, or has a complaint ticket (confirmed at the Thursday meeting) | Operations manager |
| Customer | TMS.CUST_NAME / ticket customer name | Normalized through an alias table. Branch offices are separate entities | Operations manager |
3. One-page plan
Shows you cut scope first, then set acceptance criteria the client can read, understand and sign.
- Goal: cut the dispatch lead's time per exception from 6 minutes to under 2 minutes, with root-cause classification accuracy of at least 85%.
- In scope: East China trunk-route waybills. Root causes fall into 5 categories (weather, vehicle breakdown, loading/unloading delay, address error, other).
- Out of scope: automatic replies to customers, cross-region waybills, and compensation calculation. These are for phase two.
- Deployment: one GPU server in the client's data center, running an open-source model locally. No public cloud API calls.
- Acceptance: an evaluation set of 120 historical exception waybills with human-verified causes. The dispatch lead takes part in labeling.
4. Thin slice
Shows something usable in week one, rather than a first demo in week six.
Live on day 5. Every morning at 8:00 it reads the previous day's GPS export and TMS snapshot, and produces a table of “exception waybill + aligned timeline across three systems + suggested cause + evidence.” The dispatch lead clicks “Agree / change to…” in the table. There is no chat interface and no agent, just one table and one button. The lead was using it on day one.
5. Evaluation report
Shows the results were measured, and that you know where the system still goes wrong.
Evaluation set of 120 cases, week 2 → week 4:
- Root-cause classification accuracy went from 71% to 88%. Most of the gain came from adding the rule for loading/unloading delay, based on warehouse dispatch time.
- 14 errors remain. 9 are address errors classified as “other,” because the address correction in the ticket is written in a free-text note and not structured.
- P95 latency for a single inference is 3.8 seconds. The daily batch of 200 waybills finishes in about 11 minutes, before 8:00.
6. Response to a mid-project change
Shows that when requirements changed, you neither carried on regardless nor threw out the whole plan.
In week 3, the operations manager asked to include “customer complaints” in the definition of exceptions. Response: same-day email confirming the new definition and its effective date. Impact assessment: daily exception volume is expected to rise from 180 to 240 waybills, and the batch still fits in its time window. Added 30 complaint-type samples to the evaluation set. Informed the dispatch lead and IT director. Retrospective note: “definitions will change again, so the exception rule is now a configuration setting.”
7. Runbook
Shows the client's team can keep the system running after you leave.
- Table missing at 8:00: first check the scheduled job log, then whether the GPS export reached the shared folder, then whether the file encoding changed (this once went from GBK to UTF-8).
- Suggested causes look clearly wrong: check whether the model service loaded an old prompt version after a restart. The rollback steps are in the appendix.
- Escalation contacts: the frontline dispatch lead, then the IT director, then the delivery engineer (30-day handover period).
8. Sample scoring report
Scored 0–4 on each of the six dimensions of the FDE Competency Standard. The sample deliberately keeps its weak spots: even a good project rarely scores full marks.
D1Business insight and problem framing3 / 4
Identified customer complaint response time as the real metric, and handed the definition of “exception” to the operations manager to decide. Did not quantify the cost of the current manual work.
D2Solution design and scope management3 / 4
Scope is clearly cut, with phase-two items listed. Acceptance criteria are testable. However, it does not say who decides when evaluation labels disagree.
D3Engineering implementation3 / 4
The batch pipeline is complete. The three-system timeline handles time zones and voided waybills. Making the exception definition a configuration setting is a plus.
D4Evaluation and quality assurance4 / 4
The evaluation set was labeled with the user's participation. The report gives error causes and next improvements, and the latency data is complete.
D5Delivery, security and compliance3 / 4
Deployed fully on the internal network, and the runbook can be executed. Changes were answered promptly and communicated well. Missing resource monitoring and alerts for the model service.
D6Communication and proving value2 / 4
Communication during the project was good, but the final report showed only accuracy. It did not convert the results into hours saved for the dispatch lead or the change in complaint response time.
Want to try it yourself? Pick a scenario in Simulation Projects, or first complete labs L8 and L9 in the course.