Ibtisam
Ahmed Khan

A publication on materials informatics and the trustworthiness of scientific data, written and built end to end. It hosts eighteen interactive material dashboards, an article archive on AI for materials, a resources directory, and a Lab section presenting original benchmark work with its methods and limitations stated.

An audit of whether two major open materials databases agree about the same material. Entries are matched by crystal structure rather than by chemical formula, because the same formula can hide polymorphs whose properties differ by several electronvolts. Across 344 structure-matched pairs, only 26 percent of formation energies agree within the conventional threshold. The benchmark attributes each disagreement to a specific methodological cause, and a live audit tool runs the same comparison on any composition. No language model computes any number: the deterministic core produces every value and a guard traces each figure in the explanation back to it.

Most materials models run forwards: give them a composition, they predict a property. This one runs backwards. A property surrogate is trained on real measured steel data, then a conditional generative model proposes new compositions for a target strength. Ask for a higher target and the generated alloys genuinely get stronger.

An annotated build of a Model Context Protocol server exposing the Materials Project as tools an AI can call: a bare API call, the same call wrapped as an MCP tool, then an agent that discovers and chains the tools itself. Written to understand the mechanics from the inside rather than to replace the official server.

A data pipeline that ingests live data from the Materials Project, PubChem, and ChEMBL, with NIST conventions for units. It standardises every source onto one schema, validates it, classifies records into an ontology exported to RDF and OWL, and resolves the same substance across sources. It loads the result into a SQL store and a graph view with provenance on every record, and trains a cross-validated property model. Cached, tested, and reproducible.

A composable pipeline engine where reusable components are chained by a declarative config, the same paradigm as BIOVIA Pipeline Pilot. It ingests raw lab data, harmonises mixed units from first principles, runs quality control, and emits a FAIR dataset with a data-quality report.

A research-grade study: a curated database of 55 engineering materials validated against an ontology and exported to RDF, Turtle, and OWL, exploratory analysis, a machine-learning model predicting Young's modulus (linear model, R squared 0.93, cross-validated), and an Ashby material-selection engine with real case studies.

Eighteen published material studies, one at a time, covering rare earths, lithium, cobalt, graphite, nickel, manganese, silicon, copper, steel, aluminium, titanium, magnesium, tungsten, zinc, tin, silver, gold and uranium. Each pairs a reproducible Python pipeline, running from cited sources to documented CSVs, with an interactive dashboard covering production, supply concentration by producing country, end uses, prices and physical scale. All eighteen are now hosted on my own domain.

Forensic, animated breakdowns of famous engineering failures. The first episode rebuilds the 1940 Tacoma Narrows Bridge collapse as a rendered animation, with the data and physics of aeroelastic flutter, the design flaw, and what it changed in engineering.

End-to-end pipeline that ingests PubMed scientific data, indexes it for vector search with FAISS, and generates answers through the Claude API behind a Gradio interface. A working example of scientific-data ETL over a real research database.

End-to-end sales reporting solution using a star schema data model with 20+ DAX measures including YoY growth, YTD cumulative totals, and rolling averages. Published to Power BI Service with drill-through navigation.

Weekly executive dashboard across store, department, and date dimensions. YoY index measures (Current/Previous × 100), YTD comparisons, manager commission table (0.5% of weekly sales), and mobile-optimised layout.

4-page interactive dashboard covering executive KPIs, sub-category profitability and margin analysis, customer LTV and RFM segmentation, and monthly revenue trends with new vs returning customer breakdown.

11 progressive SQL queries on Google BigQuery's public ecommerce dataset. Covers LTV segmentation, cohort retention, category margin analysis, demographic spend, traffic source performance, using CTEs, LAG, RANK, PERCENT_RANK, NTILE, rolling averages, and four-table JOINs.

Multi-table relational hospital database covering patients, appointments, departments, doctors, and billing. Applied ROW_NUMBER, RANK, CTEs, and CASE WHEN logic to produce operational dashboards and management KPI reports.

Real A/B test analysed end to end, data quality checks (mismatch removal, deduplication), sample size validation, two-proportion z-test, confidence interval, novelty effect detection via daily trend analysis, peeking problem simulation, and segment breakdown by day of week.

Full RFM segmentation model on 4,000+ UK online retail customers, 8 named segments (Champion, Loyal, At Risk, Lost, etc.) with a plain-language marketing action plan per segment. Cohort retention heatmap showing Month 1 retention averaging 3-7%.

Joined 8 relational CSV tables (100K+ orders) into a master analytical dataset. Analysed order funnel drop-off, category performance, customer LTV, delivery impact on satisfaction, and geographic revenue breakdown across Brazilian states.

Analysed Pakistan real estate listings to identify pricing drivers and investment opportunities. Applied fuzzy matching to standardise 14 city name variants, converted mixed area units, removed outliers using IQR and Z-score, and engineered new features. Produced investor-facing recommendations report.

Segmented customers across three treadmill product tiers ($1,500, $1,750, $2,500) using conditional probability tables, pairplots, correlation heatmaps, and IQR outlier detection. Identified that high-income fitness-level-5 males accounted for the majority of premium KP781 purchases.

Fine-tuned YOLOv11n on 1,026 images (4,300+ labelled instances) to detect PPE compliance on industrial sites. Achieved mAP50 of 96.6% for vest and 96.2% for helmet detection. Added persistent YOLO tracking to assign unique worker IDs and monitor compliance over time.

RAG system querying 20,000+ PubMed papers to answer complex medical questions. Automated literature fetching via Biopython, vector search via FAISS, and answer generation via Groq (Llama 3.3 70B). Deployed as a Gradio web interface.