# Project scope and implementation evidence

These notes describe the work behind the portfolio: the problem, implementation decisions, evaluation and usable output. The public studies are independent implementations with inspectable evidence. Earlier confidential client work is described separately; its original code and operating records are unavailable here.

## Support Intent: a complete NLP evaluation and review workflow

The task is to distinguish 77 banking support needs, including messages whose similar wording can lead to different handling. The implementation covers source attribution, duplicate auditing, stratified fit/validation splits, sparse text features, a trained multinomial classifier, validation-based settings and a review threshold frozen before test evaluation.

The saved result contains all 3,080 test messages, original labels, predictions, top-three scores and review decisions. The model correctly classifies 2,711 messages. Its suggestion path includes 111 incorrect suggestions; another 612 messages require closer review. The interface exposes these mistakes, category-level results and reproducible artifacts. It browses recorded predictions rather than claiming a live helpdesk integration. The engineering scope includes data preparation, model selection, error analysis, review interactions, exports and numerical verification.

## Evidence Search: comparative retrieval with traceable rankings

The task is to retrieve useful evidence when exact wording differs. The implementation compares BM25, a pretrained MiniLM encoder and reciprocal-rank fusion over 5,183 public scientific abstracts and 300 judged queries. It retains pinned sources, tokenizer/model configuration, timings, relevance judgments and every query's rankings.

The keyword method outperforms embeddings on aggregate nDCG. Hybrid search improves the recorded comparison, but it was added after inspecting the initial results and needs fresh validation. Fifty-one hybrid queries still miss every judged relevant document in their top ten. Visitors can inspect successes, misses and source excerpts. This demonstrates indexing, model integration, ranking comparison and evaluation; private document connectors and generated answers are outside the current implementation.

## Industrial AI: measurements, learned classification and reviewed explanations

Measured hydraulic recordings feed cycle comparisons and a separately trained leakage classifier. Thirty sensor features, a train-only scaler, learned coefficients and all 552 held-out predictions are retained. The 15 missed leakage cases are visible. A separate local language-model assistant selects bounded read-only evidence tools and produces a source-linked investigation draft, with observable requests, limits and human review. Generated explanations can still be wrong. The same-rig benchmark does not establish reliability on another machine.

## Confidential freelance history

Client work includes task-performing AI agents, ML-based sensor-data analytics, workflow automation for MSME businesses, and custom CRMs for organizations seeking a reliable alternative to enterprise-scale software. It also includes websites, applications, tool integrations, a residential solar configurator and manufacturing sensor-calibration/anomaly work. This breadth concerns complete operating workflows and usable business software, rather than isolated model scores.

Paid OpenAI and Google Gemini services were used for client AI delivery. Client names and engagement details remain private. The public studies above are independent technical evidence, not the paid projects themselves. They make implementation choices inspectable without recreating unavailable client code or records. Their benchmark results do not establish historical client deployment metrics, customer savings or unverified technical details.