Student Placement Prediction - MLOps Project

Servei de classificacio binaria per predir si un estudiant sera placed o not placed, amb pipeline de dades, entrenament versionat, API FastAPI, persistencia de prediccions, containeritzacio Docker, quality gates de testing i validacio CI/CD en cada push.

Dataset

Atribut	Detall
Nom	Student Placement Dataset
Font	Kaggle - sonalshinde123
URL	https://www.kaggle.com/datasets/sonalshinde123/student-placement-dataset
Mida	50.000 registres sintetics (45.000 train / 5.000 test)
Problema	Classificacio binaria: predir placement
Llicencia	CC0 Public Domain

Distribucio target aproximada: 36.2% Placed i 63.8% Not Placed.

Features i target

Columna original	Columna normalitzada	Tipus	Rang
Age	age	int	18-24
Gender	gender	int	0=Female, 1=Male
Degree	degree	int	0=B.Tech, 1=BCA, 2=MCA, 3=B.Sc
Branch	branch	int	0=ECE, 1=ME, 2=Civil, 3=CSE, 4=IT
CGPA	cgpa	float	4.5-9.8
Internships	internships	int	0-3
Projects	projects	int	1-6
Coding_Skills	coding_skills	int	1-10
Communication_Skills	communication_skills	int	1-10
Aptitude_Test_Score	aptitude_test_score	int	35-100
Soft_Skills_Rating	soft_skills_rating	int	1-10
Certifications	certifications	int	0-3
Backlogs	backlogs	int	0-3
Placement_Status	target	int	0=Not Placed, 1=Placed

Preprocessament

• Esborra Student_ID (metadada sense valor predictiu).
• Renomena columnes a format estandard en minuscules.
• Codifica categoriques dins del pipeline:
  • gender: Female=0, Male=1
  • degree: B.Tech=0, BCA=1, MCA=2, B.Sc=3
  • branch: ECE=0, ME=1, Civil=2, CSE=3, IT=4
• Converteix target a binari:
  • Placed=1
  • Not Placed=0

Arquitectura del projecte

• app/pipeline.py: ETL + normalitzacio + splits 60/20/20.
• app/train.py: entrenament, quality gate i versionat de model.
• app/api.py: endpoints health, predict i predictions.
• app/pred_store.py: persistencia JSONL de prediccions.
• .cicd/hooks/validate.sh: fase 1 de CI (tests dins Docker a cada push).
• .cicd/hooks/pre-deploy.sh: fase 2 de CI (quality gate de deployment_ready).
• models/: artifacts del model (.pkl).
• metadata/: metadata de versions (.json).
• data/: dades d'entrada, splits i predictions.jsonl.
• .env.production: configuracio del servidor CI comesa al repositori.

Nota important: metadata i model estan separats en directoris diferents:

• models/ nomes te fitxers model_vN.pkl
• metadata/ nomes te fitxers metadata_vN.json

Execucio rapida local

make setup
make pipeline
make train
make dev

API disponible a:

• http://localhost:8000/health
• http://localhost:8000/predict
• http://localhost:8000/predictions
• http://localhost:8000/docs

Exemple de prediccio

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"age":21,"gender":1,"degree":0,"branch":3,"cgpa":7.5,
       "internships":1,"projects":3,"coding_skills":7,
       "communication_skills":6,"aptitude_test_score":72,
       "soft_skills_rating":7,"certifications":2,"backlogs":0}'

Resposta esperada (format):

{
  "prediction": 1,
  "probability": 0.72,
  "model_version": "1.0.0"
}

Monitoratge de prediccions (Sessio 8)

Cada crida a predict guarda una entrada JSON Lines amb:

• timestamp
• input
• prediction
• probability
• model_version

Path configurable via .env:

PREDICTIONS_LOG_PATH=data/predictions.jsonl

Consulta historial:

curl "http://localhost:8000/predictions?limit=5"

Docker (Sessio 9)

Build:

make docker-build

Run amb persistencia de dades i logs:

make docker-up
make health

El servei en contenidor:

• usa usuari no root
• te healthcheck configurat
• persisteix data i logs via volums

Stop:

make docker-down

Testing i quality gates (Sessio 11)

Suite implementada amb pytest en tres blocs:

• tests/test_pipeline.py: preprocessament + schema
• tests/test_model.py: carrega de model + format de prediccio
• tests/test_api.py: endpoints + validacions + logging

Executar tests local:

make test

Executar tests dins Docker:

make docker-test

Estat actual: 12 tests passing local i 12 tests passing en contenidor.

CI/CD - Validacio automatica (Sessio 12)

Sessio 12 separa clarament validacio i desplegament:

• git push: executa validacio (fases 1 i 2), sense deploy.
• git tag: disparara deploy a la Sessio 13 (fases 3 i 4).

Fitxers afegits a la sessio:

• .env.production
• .cicd/hooks/validate.sh
• .cicd/hooks/pre-deploy.sh

Fase 1 (validate.sh):

• copia .env.production a .env
• executa make docker-test
• falla si qualsevol test falla

Fase 2 (pre-deploy.sh):

• llegeix METADATA_PATH
• valida que el fitxer metadata existeix
• extreu version, f1_score i deployment_ready amb jq
• bloqueja si deployment_ready no es true

Verificacio local recomanada abans de push:

bash .cicd/hooks/validate.sh
bash .cicd/hooks/pre-deploy.sh

Sortida esperada:

• Validation passed al final de fase 1
• Pre-deploy checks passed al final de fase 2

Requisits tecnics de Sessio 12:

• jq disponible per parsejar JSON en shell scripts
• Makefile amb targets docker-test, docker-up, docker-down, docker-build, health i predict
• metadata amb camp deployment_ready a l'arrel del JSON

Configuracio de servidor CI usada al projecte:

MODEL_PATH=models/model_v6.pkl
METADATA_PATH=metadata/metadata_v6.json
PORT=8085
HOST=0.0.0.0
LOG_LEVEL=INFO
LOG_FILE=logs/api.log
APP_NAME="Student Placement Prediction API"
APP_VERSION=1.0.0
PREDICTIONS_LOG_PATH=data/predictions.jsonl

Primer push a CI (sense deploy):

git push origin master

Si el servidor completa fase 1 i fase 2, el codi queda validat i llest per la Sessio 13.

Model

• Algoritme: Logistic Regression (max_iter=2000)
• Preprocessing: StandardScaler
• Accuracy interna reportada durant training: ~86.67%
• Artifact principal: models/model_vN.pkl
• Metadata versionada: metadata/metadata_vN.json

Variables d'entorn clau

MODEL_PATH=models/model_v1.pkl
METADATA_PATH=metadata/metadata_v1.json
PREDICTIONS_LOG_PATH=data/predictions.jsonl
PORT=8000
HOST=0.0.0.0
LOG_LEVEL=INFO
LOG_FILE=logs/api.log
APP_NAME=Student Placement Prediction API
APP_VERSION=1.0.0