Windows GUI
Accepts annual reports or ZIP batches and runs text processing with sentence-level sentiment analysis.
Start with one company, extend batches and fields, then resolve dependencies, startup, and handoff on the client’s computer.
The client needed topic-specific sentences and sentiment from several listed companies’ annual reports for further manual research. They specified the task, not the technical solution. I translated that goal into inputs, processing, outputs, and acceptance criteria.
Usage scenario: the client supplies reports or a ZIP, runs the tool, exports Excel / CSV, and reviews the results alongside the source material. This task sequence shaped the product.
Start with one company to validate text processing, topic filtering, sentiment classification, and export. Defer batching to reduce the number of requirements being resolved at once.
Actual usage needs introduced multi-company processing and output-field requirements. I added report / ZIP batches and adjusted XLSX column names and aggregation definitions. Batching also introduced runtime and hardware constraints that needed delivery support.
Missing dependencies and a GUI startup failure on the client’s computer exposed an incomplete delivery environment. I repackaged the runtime and supplemented the demo and guide, then verified client processing and export.
This account paraphrases project and delivery records; it does not quote client messages.
Accepts annual reports or ZIP batches and runs text processing with sentence-level sentiment analysis.
Produces structured data for further analysis, with export fields adjusted to the client's requirements.
Documents setup and operation so the client can use the tool in their own environment.
This case is anonymized. Client materials, software screenshots, and acceptance correspondence are not public. Deliverables are described in text; illustrations are not presented as runtime evidence.
The client followed the guide to complete batch processing and structured exports, then accepted the delivery and paid.
This supports the claim of a usable delivery. No quantitative sentiment accuracy, time savings, or investment returns are established here.
The tool provides information for human analysis; investment decisions remain with the user.
Acceptance focused on operation, batch processing, and obtaining outputs. This established delivery usability, but did not formalize output-quality criteria such as missed critical sentences.
If repeating the project, I would separate operational and output-quality acceptance before the first version: identify consequential misses, then fix labeled samples, per-class metrics, and version records to guide model selection and retesting.
This is a retrospective improvement plan, not a newly completed client acceptance cycle. Original test materials must be recovered before reporting verifiable classification performance.
I owned scope, processing flow, sentence-level model and text-processing design, the Windows interface, export fields, and handoff support. The client accepted the delivery through actual use. Client originals are private; interviews can explore requirements, iterations, and environment fixes.