AI & ML / business application

AI work is useful when it connects to a business decision.

These are training projects from the UT Austin / Texas McCombs AI & ML program. They are not client case studies. They show the judgment that matters for business AI work: problem framing, source data, model choice, metrics, recommendation, deployment path, and human review boundary.

forecasting system

model

RF tuned

0.932 test set

MAPE

3.8% business view

deploy

HF Docker Spaces

sales forecasting / model to API to UI

Lead artifact / SuperKart

The strongest example moves from model to API to UI.

SuperKart is relevant because it does not stop at a notebook. It compares models, selects a tuned Random Forest, serializes the pipeline, exposes a Flask API, wraps it with Streamlit, and ships through Docker-based Hugging Face Spaces. That is the kind of path a business can understand: data, model, interface, and recommendation.

Backend Space Frontend Space

ship record

$ train tuned Random Forest pipeline
$ serialize superkart_sales_model.joblib
$ serve predictions with Flask
$ wrap inputs in Streamlit
$ deploy backend + frontend spaces

Selective archive

The range matters because each business problem has different tradeoffs.

These are sanitized summaries. They stay here as evidence of judgment across problem types, not as a promise that every business needs AI or that a model should replace a human decision.

retrieval system

Medical RAG assistant

chunk medical manuals

embed with sentence transformers

retrieve top context

judge grounding

sanitized project

Medical RAG assistant

A healthcare question-answering prototype grounded in medical manuals rather than standalone model memory.

pressure: A model answer is risky when it is detached from source material and review.
installed: Chunked manuals, embeddings, retrieval, answer comparison, and grounding checks.
evidence: RAG notebook artifact with retrieved context and relevance judgments.
handoff: AI answers need source material, safeguards, and human review.

neural network

test recall

0.837 failure class

val recall

0.919 selection signal

metric

F1 tradeoff view

stack

Keras neural network

recall-first model

ReneWind maintenance

A failure-prediction project where missing a failure mattered more than producing a tidy accuracy score.

pressure: Unplanned generator failure is the expensive miss, so tidy accuracy is not enough.
installed: Neural-network comparison with recall-first evaluation and false-alarm tradeoff review.
evidence: Final selected model reached 0.837 test recall for failures.
handoff: Maintenance recommendations are framed as prioritization, not blind automation.

decision support

01prioritize cases

02flag borderline files

03monitor geography bias

04retrain with policy drift

responsible framing

EasyVisa classification

A visa certification model framed as prioritization support, not automated decision-making.

pressure: Classification support can become irresponsible if it hides imbalance, bias, or review boundaries.
installed: F1 and ROC-AUC evaluation with imbalance notes and human-review framing.
evidence: Runbook-style recommendations for prioritization, fairness monitoring, and retraining.
handoff: Keep borderline and policy-sensitive cases in human hands.

marketing model

ROC-AUC

~0.95 campaign scoring

signal

CD relationship depth

segment

>$100K income threshold

use

target not spam

segmentation signal

AllLife Bank targeting

A personal-loan targeting project focused on customer segments, model interpretation, and outreach economics.

pressure: Marketing outreach wastes money when targeting ignores segment economics and customer fit.
installed: Model comparison, interpretation, segment review, and business-impact framing.
evidence: Around 0.95 ROC-AUC with income, education, CD accounts, and card spend as signals.
handoff: Use targeting to prioritize useful outreach, not spam a list.

computer vision

helmnet-safety-vision

load image classestrain CNN baselinecompare VGG16 headreview metric tableclean final claim

visual model exploration

HelmNet safety vision

A helmet-detection exploration using CNN and VGG16 transfer-learning paths for safety monitoring.

pressure: Computer-vision safety claims need clean metrics before they become public claims.
installed: Scratch CNN, VGG16 head, and data-augmentation comparisons.
evidence: Visual model artifact presented as exploration until the final claim is cleaned.
handoff: Do not publish a best-model claim until the metric table and conclusion agree.

Model handoff

The useful pattern is problem, model, metric, ship, handoff.

This is the bridge into consulting: AI only helps when it has source material, evaluation, deployment thinking, privacy boundaries, and human ownership.

problem

Start with business use

Forecast revenue, prioritize outreach, flag risk, answer from manuals, or detect safety state.

business context

model

Choose the right shape

Regression, classification, neural network, computer vision, and RAG solve different problems.

model comparison

metric

Make the tradeoff visible

RMSE, MAPE, F1, recall, ROC-AUC, and grounding scores tell different stories.

evaluation table

ship

Move toward operation

The best project connected model output to API, UI, deployment, or clear business recommendation.

API / UI / recommendation

handoff

Keep humans in control

The responsible pattern is decision support, monitoring, retraining, and clear ownership.

runbook logic