Architecture Decision Records

Alternatives considered · 5

• TF-IDF onlyFast and credible as a baseline, but insufficient to demonstrate Hugging Face/PyTorch fine-tuning.
• Frozen sentence embeddings plus logistic regressionLower implementation risk and valid as a fallback, but demonstrates less end-to-end transformer training.
• BERT-base, RoBERTa or DeBERTaPotentially stronger but larger/slower. Additional performance is not required to prove the engineering point.
• Small generative LLMWould make classification slower, harder to calibrate, and harder to evaluate without adding business value.
• TensorFlowTechnically valid, but the user’s local workflow and target portfolio benefit more from a straightforward PyTorch/Hugging Face path. Suppor…

Reconsideration trigger

Reconsider only if: - the checkpoint cannot load or make a forward pass on the target environment; - fine-tuning cannot produce a valid artifact by Hour 9 after the documented fallback; - measured latency or memory exceeds the hardware budget; - the target role explicitly requir…

Alternatives considered · 4

• CLINC150Offers explicit out-of-scope examples and more domains, but broadens the business context and adds taxonomy complexity.
• Proprietary or scraped support ticketsWould be more realistic but introduces privacy, licence, cleaning and annotation problems that do not fit the weekend.
• Fully synthetic datasetUseful as a fallback and for tests, but not credible as the primary source for portfolio performance claims.
• Multiple support datasetsCould improve domain-shift analysis, but violates the one-primary-dataset constraint and increases normalization work.

Reconsideration trigger

Reconsider if: - the pinned revision or licence becomes unavailable or ambiguous; - the loader cannot reproduce the upstream split; - the target job is explicitly non-NLP or requires another domain; - a real, legally usable employer dataset is later supplied. Changing the primar…

Alternatives considered · 4

• CLI onlySimplest implementation, but shows less boundary validation and is less representative of how a model becomes part of a system.
• CLI plus FastAPIConvenient but duplicates inference presentation and testing. Training/eval commands already provide sufficient CLI operation.
• Gradio or StreamlitMore visual, but a frontend would consume time without improving evaluation or engineering evidence.
• Multiple servicesUnjustified for one model and one local process.

Reconsideration trigger

Reconsider if: - API implementation threatens the Sunday Hour 13 gate; - the target environment explicitly prohibits HTTP serving; - a later consumer requires batch-file inference or a specific protocol. A future batch interface must reuse the predictor module and receive its ow…

Alternatives considered · 5

• `requirements.txt` and virtualenvSimple, but weaker lock resolution and grouped development dependencies.
• PoetryValid, but adds no advantage over the selected lightweight workflow.
• DockerfileUseful for deployment, but GPU pass-through and model-weight handling increase weekend risk. A CPU inference image would not validate the t…
• Docker ComposeNo second service exists, so Compose would be decorative.
• Model registry or object storageUnnecessary for one local artifact.

Reconsideration trigger

Reconsider after the MVP when: - all MUST evaluation and tests pass; - a deployment target is known; - a repeatable CPU or NVIDIA container can be validated without embedding model weights or secrets; - an interviewer or employer specifically requests container packaging.

Reconsideration trigger

Revisit if the platforms adopt a common repository skill location or either platform removes the discovery behavior on which an adapter depends.