Taiwo Alabi

Progressed from Data Analyst Intern; the team extended the placement. Worked inside a network that carries 400 million call records a day, which is where I learned what scale does to a mistake.

• Data governance & migration, Supported customer data migration with the solutions architect; classified records by type and sensitivity for auditable migration. Designed 440+ UAT cases with the test team, validating migration integrity and data controls before cutover. Documented data handling rules and GDPR-aligned data cataloguing for migrated datasets.
• Data quality & analytics, Profiled, cleaned and validated 100k+ user engagement rows in SQL; found and fixed a 14% MAU reporting discrepancy. Gathered stakeholder requirements and turned them into governed, reusable datasets and measures. Owned Power BI reports, dashboards, workspaces and datasets; added advanced analytics for self-service.
• Reporting & automation, Built SSRS and Tableau dashboards tracking 30+ KPIs across departmental operations. Automated daily CRM data extraction in Python; saved ~15 hours of manual reporting weekly. Designed database tables from 20+ business requirements; supported rollout of new software systems.
• Process modelling, Modelled the customer-care triage process in BPMN 2.0 before writing any code, so the implementation matched a diagram a non-engineer could review. The same BPMN scaffold is what the automated triage bot (see Projects) was built on.

Eighty hours of contract work, remote. Evaluated LLM outputs against structured rubrics for factual accuracy, safety, and hallucination. Wrote deliberately tricky prompts to find where the rubric broke down; documented the decision rules so a second annotator would agree.

Translated raw data into actionable business insights by engaging with stakeholders to define reporting requirements. Managed Power BI assets including reports, dashboards, workspaces and underlying datasets. Built scalable data models and integrated advanced analytics into reporting pipelines. Led data profiling, cleaning and transformation workflows. Internship extended by the team due to strong performance.

The customer-care process at a mobile-network operator routes complaints about SIM, billing and network issues to three different back-end teams. I modelled the as-is process in BPMN 2.0, automated it with a TF-IDF + Naive Bayes classifier, and gated the classifier behind a confidence threshold so uncertain tickets escalate to a human instead of being forced into the wrong bucket.

• 95.2% routing accuracy on held-out tickets, the held-out set is the right measure, not training accuracy.
• 0.55 confidence gate, tickets below this score are escalated, not guessed. The threshold was tuned on validation, not chosen by default.
• BPMN first, code second, the implementation matched a diagram a non-engineer could review before any line of Python was written.
• Second-bot validation, a separate process validates each routing decision against the operator's CRM, catches drift, writes the daily report. The two manual processes it replaced (intake log + dispatch log) are auditable from the bot's output.

Why I list this: this is the project that has the strongest case for a CSEP role. It was the live work, the BPMN is the proof, and the escalation gate is the rule that says "I know when I am not sure."

Retrieval system over patient notes: surfaces critical history (e.g., prior allergies, contraindications) in seconds. The interesting design constraint was not accuracy, it was scoping, wrong-patient records have to be excluded BEFORE ranking, not filtered after, or the model will surface the wrong patient's allergy list and the clinician will see it first.

• Schema-level scoping, patient identity is enforced at the database layer, so even a misconfigured query cannot leak across patients.
• 30 adversarial probes, zero leaks, the test set was designed to break the model, not pass it.
• On-premise LLM, data never leaves the hospital network. The PII is stripped before inference. There is a full audit log on every access.
• Four Docker services, LangGraph orchestrates retrieval, ranking, generation, and the audit log. Each service is independently deployable.

Flags risky clauses in privacy policies. Grounded in a 102-clause reference set (52 fair, 50 unfair). Every flag carries the exact clause text it was drawn from, so the answer is "this clause, this law, this is why", not "the model said so."

• 87% F1 on the sealed test set, with 0.78 precision and 0.97 recall.
• Latency cut from 28 s to under 10 s by re-ranking with a smaller cross-encoder only on the top-k candidates from FAISS.
• 4,527-chunk retrieval index of GDPR, EU consumer law, and Irish Data Protection Commission guidance, updated as the regulations update.
• Up to 15 flagged clauses per analysis are passed to Claude for a second-pass verification. The final answer is the conjunction of the two.

A similarity gate that blocks non-compliant AI-generated product imagery before it reaches a client. The interesting bit is the threshold: it was tuned on validation data so the model never once raised a false alarm, because a false alarm was the error that would have reached a client.

• F1 0.90, precision 1.00 on the sealed test pairs.
• Baselines measured first: pixel difference F1 0.59, kNN F1 0.26. The gap between 0.26, 0.59 and 0.90 is the result.
• Acceptance cutoff moved off the model's default 0.5 to 0.4821. That one change recovered all 8 misses and created no new false positives. The default 0.5 was not a fact, it was a habit.

Focus: AI evaluation and critique, evaluation design, deployment patterns, sliced evaluation, confidence calibration.

Modules covered: applied machine learning, deep learning, natural language processing, knowledge representation, data governance, ethics of AI. Practicum modules on the Master's include the IAPA module that produced the Customer Care Process Automation (see Projects).

Final-year project: a process-modelled complaint-routing proof of concept. That work became the seed of the Customer Care Process Automation I built at Huawei.