Case-study library
Analyses, dashboards, and AI prototypes evaluated by the decisions they support, the evidence they expose, and the limits they state, not by a list of tools.
Every figure in these case studies that was measured against a baseline. The units differ and are not comparable to each other; the only thing all 20 share is which way the result came out.
| Project | Compared | Figure | Unit | Direction |
|---|---|---|---|---|
| arXiv Recommender | Hybrid blend against TF-IDF | 0.199 | MAP@10 | For |
| arXiv Recommender | Hybrid over TF-IDF on MAP@10 | 1.34× | Ratio | For |
| arXiv Recommender | MiniLM neural tower against TF-IDF | 0.121 vs 0.149 | MAP@10 | Against |
| Automotive Analyst | Adversarial SQL suite against the guardrail | 43 / 43 | Attacks blocked | For |
| Automotive Analyst | Legitimate analytical queries against the guardrail | 16 / 16 | Queries accepted | For |
| Automotive Analyst | Quoted identifiers against my own allow-list | 1 bypass | Bypasses found | Against |
| Black Box AI | Hybrid retrieval against semantic retrieval | 0.902 vs 0.853 | MRR | No separation |
| Dispatch Optimizer | Greedy plus 2-opt against the naive baseline | 12 fewer | SLA breaches per day | For |
| Dispatch Optimizer | Greedy plus 2-opt against the naive baseline | 3.35 hrs | Overtime per day | For |
| Dispatch Optimizer | CP-SAT at the shipped 8s budget against greedy plus 2-opt | 5 behind | SLA breaches per day | Against |
| Dispatch Optimizer | The naive baseline measured against itself | 45.6% | Deadlines missed | Against |
| Grid Intelligence | Model against the published EIA day-ahead forecast | 5.49% vs 8.30% | MAPE | For |
| Grid Intelligence | Model against the published EIA day-ahead forecast | +10.0% | RMSE skill | For |
| Grid Intelligence | Model against EIA, counted per authority | 35 / 48 | Balancing authorities | For |
| Manufacturing Intelligence Platform | Injected process events against what the pipeline recovered | Round trip | Signal recovery | For |
| Manufacturing Intelligence Platform | Propagation join against assuming the origin station | 69% | Defects downstream | For |
| Manufacturing Intelligence Platform | Rediscovered step change against the seeded one | 45% | Defect reduction | For |
| Water System Risk & Funding Priority Index | Weighted index against counting prior violations | 0.740 vs 0.734 | ROC AUC | Against |
| Water System Risk & Funding Priority Index | Compliance component alone against the full index | 0.781 | ROC AUC | Against |
| Water System Risk & Funding Priority Index | Index stripped of compliance against chance | 0.447 | ROC AUC | Against |
Where to start
Every demo linked below is live and every repository is public, so none of this has to be taken on trust.
8 projects shown
Natural-language aviation safety analysis over NTSB final reports, combining guarded SQL, cited narrative retrieval, validated charts, visible queries, and audit trails.
Value: answers questions over 7,462 NTSB reports while keeping exact SQL counts and cited narrative evidence separate and auditable.
NTSB public data with source citations, SQL evidence, and an exploratory labeled evaluation.
Constraint-based field-service dispatch optimizer for the synthetic Atlas Field Services. An OR-Tools CP-SAT vehicle-routing-with-time-windows plus assignment model turns a backlog, technician capacity, skill matching, SLA rules, and travel into feasible job assignments, benchmarked live against a manual dispatch baseline.
Value: the project that does not just report on a business but makes a decision, benchmarked across 200 randomized scenarios against a greedy dispatcher with a 2-opt improvement pass rather than against a strawman. The comparison surfaced its own finding: the shipped eight-second solver budget sits below the point where constraint programming beats a good heuristic.
Synthetic benchmark with a CP-SAT solver and two named planning baselines.
Hypothesis testing and statistical inference on CFPB HMDA mortgage data, with pre-specified tests, sensitivity analyses, multiplicity controls, effect sizes, confidence intervals, and explicit causal caveats.
Value: turns 210,643 HMDA records into a screening dashboard where findings are checked across parametric, non-parametric, resampling, and Bayesian methods.
Public mortgage data with multi-method validation and explicit causal limits.
Factory analysis using a reproducible synthetic assembly dataset, PostgreSQL star schema, validation checks, and an executive decision dashboard.
Value: turns raw operational logs into a clear plant-health view, with a defensible path from source records to OEE, yield, defect, and downtime findings.
Synthetic process data with validation evidence, current screenshots, and source code.
Bring-your-own-key text-to-SQL agent for the synthetic factory warehouse, with client-side model calls, schema grounding, read-only SQL guardrails, and visible query evidence.
Value: lets anyone query the warehouse in plain English with zero LLM keys held on the backend and read-only SQL guardrails on every generated query.
Synthetic manufacturing data with SQL guardrails, current screenshots, and source code.
Explainable Ohio public drinking water screening model using public data, transparent scoring, service-area boundaries, source-water overlays, API-backed search, and dashboard delivery.
Value: converts 16,339 Ohio water-system records into a transparent screening tool, then backtests it honestly: the weighted index does not beat simply counting prior violations, and the case study says so and explains why.
Ohio public data with 22 validation checks, current screenshots, and methodology notes.
US and European electricity analysis using scheduled public-data ingestion, time-series modeling, anomaly scoring, forecasting, source freshness, and dashboard delivery.
Value: turns scattered grid feeds into one view of demand, forecasts, and anomalies across 65+ US balancing authorities and 20 European zones.
Public energy data with current screenshots, forecasting detail, and architecture notes.
Hybrid arXiv paper recommender on a 28,000-paper OpenAlex snapshot of Computer Science papers. Compares popularity, TF-IDF, a MiniLM sentence-transformer, citation-graph ALS, and a hand-tuned linear blend, with bootstrap-CI evaluation, exact FAISS inner-product serving, and a side-by-side interactive dashboard.
Value: the hybrid blend reaches MAP@10 0.199 against 0.149 for TF-IDF on 2,000 held-out seeds, with non-overlapping bootstrap intervals. The benchmark also shows the MiniLM neural tower losing to plain TF-IDF, which is the more useful finding.
Public research data with bootstrap intervals, a held-out metric leaderboard, and source code.
