Mojammel
Hossain

At Presidency International School, I was brought in to extract and migrate data from a vendor-locked ASP.NET system. No source code access — only the admin dashboard and data exports. Applied QA techniques to validate every record from the outside: cross-referenced schemas by observing UI behavior, traced data relationships across screens, detected anomalies through export comparisons. That experience of sanitizing output from a system I couldn't modify became this CLI + CI action. 5 deterministic validators now do the same for AI-generated code. Same philosophy: validate before you trust.

The Question Bank LLM pipeline generates new question variations based on past paper structures — teachers can produce multiple exam variants from the same syllabus coverage. Claude API handles the generation; the eval framework verifies every output. 5 deterministic scorers: exact match, fuzzy match, hallucination detection, format validation, semantic. JSON-defined specs, CI-ready. No LLM-judging-LLM circularity. Built because teachers trusting AI-generated exam content needed a verification layer with teeth.

RAG systems degrade silently — embeddings drift, retrieval misses context, LLMs hallucinate over bad chunks. This evaluation harness measures retrieval quality (Precision@k, Recall@k, MRR, NDCG), faithfulness, hallucination rates, and context relevance. Every evaluator is deterministic and CI-ready. The philosophy scales: if you can validate LLM output for exam papers, you can validate it for any retrieval-augmented system.

Production performance review system for ~300 teaching staff. Role-based evaluation sessions with multi-part weighted scoring, audit logging, and printable reports. First evaluation cycle completed with zero data issues. Built while simultaneously extracting, sanitizing, and validating data from the school's locked legacy .NET system for migration.

Architected a form-builder-driven KPI evaluation engine where organisations configure their own weighted, multi-form reviews. Role-based access, multi-evaluator aggregation, scoring across arbitrary form structures. Built solo end-to-end — from the MongoDB schema to the Playwright suites that keep it honest.