GitHub Presentation

Recruiter two-minute path#

A recruiter should see, in this order:

one-sentence problem and portfolio purpose;
honest implementation status;
measured baseline-versus-transformer table;
one architecture diagram;
quick-start and one API example;
tests and reproducibility evidence;
limitations and non-goals.

Suggested repository description#

Local-first, confidence-aware banking intent classifier comparing TF-IDF with a fine-tuned DistilBERT model through reproducible evaluation and a tested FastAPI boundary.

Final README structure#

Title and one-sentence outcome.
Honest status badge/table.
Why this project exists.
What is implemented.
Results.
Architecture.
Quick start.
API example.
Evaluation methodology.
Tests and reproducibility.
Limitations.
Repository structure.
Decisions and post-weekend parking lot.
Dataset/model attribution and licence.

Quick start#

The implementation README should prefer:

git clone <repository-url>
cd intentguard
make setup
make data baseline train evaluate test
make serve

If transformer training is too long for a recruiter, document a second demo path that downloads a release artifact only after an artifact has actually been published. Do not include a dead or planned download command.

Sample input and output#

Input:

{"text":"How do I activate my new card?"}

Output schema:

{
  "request_id": "req_...",
  "decision": "accept",
  "intent": "activate_my_card",
  "confidence": "<measured per request>",
  "threshold": "<selected from validation>",
  "input_truncated": "<true when tokens were removed>",
  "model_version": "<artifact version>",
  "latency_ms": "<measured locally>"
}

Evaluation table template#

Replace every placeholder from generated reports. Never estimate these values.

Model	Test macro-F1	Accuracy	ECE	p50 latency	p95 latency
TF-IDF + logistic regression	Not measured	Not measured	Not measured	Not measured	Not measured
Fine-tuned DistilBERT	Not measured	Not measured	Not measured	Not measured	Not measured

Selective-prediction table:

Threshold source	Coverage	Accepted accuracy	Selective risk	Abstention rate
Validation-only policy	Not measured	Not measured	Not measured	Not measured

Implemented versus planned status#

Use:

Capability	Implemented	Measured	Notes
Validated BANKING77 pipeline	☐	☐
TF-IDF baseline	☐	☐
Fine-tuned DistilBERT	☐	☐
Validation-only threshold	☐	☐
FastAPI inference	☐	☐
CPU test suite	☐	☐
Docker	No	No	Post-weekend only
Cloud deployment	No	No	Not required

Concise architecture explanation#

Offline commands prepare BANKING77, train a lexical baseline and one DistilBERT classifier, select an abstention threshold using validation data, and save a local versioned artifact. A single FastAPI process loads that artifact and returns either an accepted intent or an abstention. Evaluation reloads saved artifacts and compares both models on the untouched test split.

Limitations block#

The first screenful or two should state:

English banking taxonomy only;
confidence is not certainty;
unsupported-query fixture is curated, not a general OOD benchmark;
no real users, production traffic, or business-impact measurement;
latency applies only to the documented local environment;
one global threshold may not treat every intent equally.

Five-minute interview demonstration#

0:00–0:40 — Problem and scope#

“This is a deliberately weekend-sized production ML project. It classifies fine-grained banking-support intents, compares a lexical baseline with DistilBERT, and abstains on low-confidence requests. It does not pretend to be a complete support platform.”

Show the status and non-goals.

0:40–1:30 — Architecture#

Show the compact diagram and explain:

upstream test isolation;
shared evaluation pipeline;
local artifact;
one serving process.

1:30–2:30 — Evaluation#

Open the generated comparison table:

identify the baseline;
explain macro-F1;
show whether the transformer actually improved;
show risk/coverage and the validation-only threshold;
mention one failure class.

2:30–3:40 — Live API#

Run:

make serve
curl -s http://127.0.0.1:8000/health
curl -s -H 'Content-Type: application/json' \
  -d '{"text":"How do I activate my new card?"}' \
  http://127.0.0.1:8000/v1/predict

Then run the prepared abstention example through make demo.

3:40–4:30 — Engineering evidence#

Show:

saved provenance;
label/threshold consistency checks;
requirement-to-test traceability;
one metric regression test;
privacy-conscious log fields.

4:30–5:00 — Limitations and decision quality#

“The abstention fixture does not prove general OOD performance. I excluded a frontend, database, Docker, cloud deployment and monitoring stack because they would not improve the core weekend evidence. The next meaningful study would be stronger OOD data or per-class threshold analysis—not more infrastructure.”