I find the data, compare the alternatives and keep working when the first explanation falls short. These investigations show how I turn unfamiliar problems into questions that can be tested and decisions that can be explained.
From manufacturing data to environmental surveys and operational handovers: nine investigations from my wider research programme. Select one to follow the question, the work and what changed.
Animated workflow illustration · figures are from archived analysis
74conflicting content identities
Manufacturing / data integrity
Before trusting the model, interrogate the data.
THE QUESTION
Could a fabric-image dataset support a defensible development decision?
MY INVESTIGATION
I investigated the data behind the proposed approach, compared content identities rather than relying on filenames, and looked for duplication and conflicting labels. I developed a split ledger that holds conflicting material apart from the provisional training and evaluation sets.
WHAT THE WORK REVEALED
The archived ledger contains 5,520 paths but 5,394 distinct content identities. Seventy-four identities have conflicting folder labels. That changes the preparation work needed before a model result can be trusted.
Animated workflow illustration · figures are from archived analysis
17 / 43crossings caught: full model / baseline
Monitoring / model evaluation
A better prediction must support a better decision.
THE QUESTION
Does the more elaborate model improve the decision an alert system needs to make?
MY INVESTIGATION
I investigated the sensor data, compared alternative models and examined their behaviour at an explicit analytical threshold. I challenged the result at the level of missed events and false alerts, rather than accepting one average score as the answer.
WHAT THE WORK REVEALED
Across 3,250 archived prediction opportunities, the full model caught 17 of 114 threshold crossings and produced 11 false alerts. The simple persistence baseline caught 43 and produced 71 false alerts. Fewer false alarms came with more missed crossings.
Animated workflow illustration · figures are from archived analysis
348matched route-and-month cells
Ecology / comparison design
Make the comparison earn its conclusion.
THE QUESTION
How much can a changing encounter rate tell us when survey effort and location also change?
MY INVESTIGATION
I investigated the survey records and compared routes, calendar months and search effort. I developed a matched comparison with fixed effort weights so the result could be examined on a more consistent basis.
WHAT THE WORK REVEALED
The retained comparison covers 29 routes and 348 shared route-month cells. Holding 2017 effort weights fixed gives a 5.53% increase in standardised encounters per kilometre from 2017 to 2018. The result needs that context; a headline total cannot supply it.
Animated workflow illustration · figures are from archived analysis
3.52kWh/m³ · weighted historical example
Infrastructure / analytical comparability
Compare the same thing before comparing performance.
THE QUESTION
Can energy figures from different periods and operating boundaries be compared fairly?
MY INVESTIGATION
I investigated the definitions behind the figures, separated electricity production from consumption and examined the effect of weighting. I used a historical worked example to expose how a seemingly reasonable average can change the answer.
WHAT THE WORK REVEALED
For the two retained historical observations, total electricity divided by total delivered volume gives 3.52 kWh/m³. An unweighted mean gives 3.77. The calculation illustrates why the denominator and boundary matter.
Animated workflow illustration · figures are from archived analysis
26.85%estimated-value fall that removes stated headroom
Resource recovery / feasibility
Work through the economics beneath the opportunity.
THE QUESTION
What remains of an apparent recovery opportunity after cost boundaries and sensitivity are examined?
MY INVESTIGATION
I investigated a historical recovery case, compared the reported value and cost accounts, and looked for excluded items and inconsistencies. I developed sensitivity checks to show where the apparent headroom disappears.
WHAT THE WORK REVEALED
The retained historical account has estimated value of $7.42m and stated costs of $5.43m: about $1.99m of unallocated headroom. A 26.85% fall in the estimated value removes that headroom before additional excluded costs.
Animated workflow illustration · figures are from archived analysis
4trust-specific review proposals
Public services / operating design
Find which part of the wait can actually be changed.
THE QUESTION
Which discharge delays concern coordination, and which require capacity that a new process cannot create?
MY INVESTIGATION
I researched the published records and developed separate review proposals for four large NHS trusts. I compared the reasons for delay and built the review around ownership, referral, acknowledgement and the evidence needed for a useful handover.
WHAT THE WORK REVEALED
The archive distinguishes hospital process, interface process and unavailable care capacity. It sets out different questions and case-sampling approaches rather than treating every reported delay as a coordination failure.
Animated workflow illustration · figures are from archived analysis
4distinct operational investigations
Utilities / evidence of completion
A closed record should tell you what was resolved.
THE QUESTION
How do reported concerns, repair records, notification and recovery connect to evidence of a resolved problem?
MY INVESTIGATION
I investigated published material for four regional water companies. I developed distinct review scopes for concern-to-action, reporting chronology, repair closure and incident recovery, and checked where existing oversight might overlap.
WHAT THE WORK REVEALED
The work separates administrative completion from physical verification and distinguishes reporting from prevention. The archive also defers one broad review offer where an existing review could overlap: the scope needs a genuine uncovered question.
Animated workflow illustration · figures are from archived analysis
3transfer assumptions: zero / half / full
Service design / commercial reasoning
Turn a promising idea into a testable business case.
THE QUESTION
Would a reminder intervention create genuinely additional attended appointments and net contribution?
MY INVESTIGATION
I investigated the evidence behind the idea and examined whether a published result could transfer to another service. I built zero, partial and full-transfer scenarios, then separated attendance from collected revenue and incremental cost.
WHAT THE WORK REVEALED
The resulting model asks what happens when an extra attendance replaces another booking, incurs staff effort or produces no additional collected revenue. A published effect becomes an input to a test, not a guaranteed forecast.
Animated workflow illustration · figures are from archived analysis
47minutes between two recorded milestones
Systems analysis / chronology
Find the time hidden between the milestones.
THE QUESTION
Are physical transfer, formal handover and departure being treated as if they were the same event?
MY INVESTIGATION
I reconstructed the chronology in a published investigation, separated the milestones and compared clock times with narrative durations. I retained discrepancies instead of silently choosing the most convenient version.
WHAT THE WORK REVEALED
The archived reconstruction distinguishes approximate physical transfer, formal handover and crew departure. One interval is 47 minutes; the recorded total also differs from a narrative duration. Those differences identify questions for the record, not automatic savings.
Keep the question open. Make the next version stronger.
The retained work includes successive public-data reviews, comparison checks and revised decision models. The examples below show where that scrutiny changed the picture: a label conflict, an error trade-off, a missing cost or a handover that needed a better definition.
My contribution is the investigation and development: finding the material, comparing the options, challenging the gaps and shaping the response. AI assists the work; the direction and critical review are mine.
01 / Manufacturing / data integrity
Fabric pilling
Before trusting the model, interrogate the data.
My investigation
I investigated the data behind the proposed approach, compared content identities rather than relying on filenames, and looked for duplication and conflicting labels. I developed a split ledger that holds conflicting material apart from the provisional training and evaluation sets.
The question
Could a fabric-image dataset support a defensible development decision?
What the work revealed
The archived ledger contains 5,520 paths but 5,394 distinct content identities. Seventy-four identities have conflicting folder labels. That changes the preparation work needed before a model result can be trusted.
WHAT THIS BRINGS TO A TEAM
Trace the source of apparent accuracy before investing in a more elaborate model.
Evidence, source record & context
Fabric identity release ledger and completed analysis, 18 September 2026. Identity counts recomputed from the retained CSV on 7 October 2026.
Exact-content checks do not resolve near duplicates, common fabric lots, label truth or image rights. This is a dataset audit, not a validated pilling predictor.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
A better prediction must support a better decision.
My investigation
I investigated the sensor data, compared alternative models and examined their behaviour at an explicit analytical threshold. I challenged the result at the level of missed events and false alerts, rather than accepting one average score as the answer.
The question
Does the more elaborate model improve the decision an alert system needs to make?
What the work revealed
Across 3,250 archived prediction opportunities, the full model caught 17 of 114 threshold crossings and produced 11 false alerts. The simple persistence baseline caught 43 and produced 71 false alerts. Fewer false alarms came with more missed crossings.
WHAT THIS BRINGS TO A TEAM
Make the operational trade-off visible, then choose the next test around the cost of each type of error.
Evidence, source record & context
Dairy ablation prediction ledger and completed decision tables, September 2026. Counts recomputed from the retained CSV on 7 October 2026.
The 39°C threshold is an analytical choice, not a veterinary diagnosis. Repeated sensor predictions are correlated and are not unique animals or disease cases. This was exploratory analysis.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
I investigated the survey records and compared routes, calendar months and search effort. I developed a matched comparison with fixed effort weights so the result could be examined on a more consistent basis.
The question
How much can a changing encounter rate tell us when survey effort and location also change?
What the work revealed
The retained comparison covers 29 routes and 348 shared route-month cells. Holding 2017 effort weights fixed gives a 5.53% increase in standardised encounters per kilometre from 2017 to 2018. The result needs that context; a headline total cannot supply it.
WHAT THIS BRINGS TO A TEAM
Test whether a persuasive trend survives a fairer comparison before attributing a cause.
Evidence, source record & context
Matched route-month ledger and completed Guam analysis, 18 September 2026. Weighted arithmetic recomputed on 7 October 2026.
Encounter rates are not independently measured population density. Matching route and month does not control every difference in observers, conditions, detection or intervention timing.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
Compare the same thing before comparing performance.
My investigation
I investigated the definitions behind the figures, separated electricity production from consumption and examined the effect of weighting. I used a historical worked example to expose how a seemingly reasonable average can change the answer.
The question
Can energy figures from different periods and operating boundaries be compared fairly?
What the work revealed
For the two retained historical observations, total electricity divided by total delivered volume gives 3.52 kWh/m³. An unweighted mean gives 3.77. The calculation illustrates why the denominator and boundary matter.
WHAT THIS BRINGS TO A TEAM
Make an engineering comparison meaningful before turning it into an efficiency or investment claim.
Evidence, source record & context
Water completed analysis, 18 September 2026; retained historical electricity and delivered-volume observations. Weighted ratio recomputed on 7 October 2026.
This is a historical comparison example used within the wider Saudi water-energy investigation. It is not a measured performance result for a Saudi plant or a current efficiency benchmark.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
Work through the economics beneath the opportunity.
My investigation
I investigated a historical recovery case, compared the reported value and cost accounts, and looked for excluded items and inconsistencies. I developed sensitivity checks to show where the apparent headroom disappears.
The question
What remains of an apparent recovery opportunity after cost boundaries and sensitivity are examined?
What the work revealed
The retained historical account has estimated value of $7.42m and stated costs of $5.43m: about $1.99m of unallocated headroom. A 26.85% fall in the estimated value removes that headroom before additional excluded costs.
WHAT THIS BRINGS TO A TEAM
Identify the assumptions that can overturn an attractive project before making a commitment.
Evidence, source record & context
STRATA completed analysis, 18 September 2026, reviewing the historical Wagner and Raymond recovery account. Value-minus-cost arithmetic rechecked on 7 October 2026.
Historical nominal US dollars; estimated value is not actual sales or profit. Offsite haulage was excluded, a royalty discrepancy remains recorded, and this does not establish the economics of a new site.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
Find which part of the wait can actually be changed.
My investigation
I researched the published records and developed separate review proposals for four large NHS trusts. I compared the reasons for delay and built the review around ownership, referral, acknowledgement and the evidence needed for a useful handover.
The question
Which discharge delays concern coordination, and which require capacity that a new process cannot create?
What the work revealed
The archive distinguishes hospital process, interface process and unavailable care capacity. It sets out different questions and case-sampling approaches rather than treating every reported delay as a coordination failure.
WHAT THIS BRINGS TO A TEAM
Define a problem narrowly enough that a team can act on it and measure whether the change helped.
Evidence, source record & context
Four hospital review records and buyer-research comparison, September 2026. Public-data summaries and proposed work packages inspected for this portfolio.
Independent research and proposals, not commissioned trust engagements or demonstrated bed-release improvements. Public repeated counts are not unique patients or automatically avoidable delays.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
A closed record should tell you what was resolved.
My investigation
I investigated published material for four regional water companies. I developed distinct review scopes for concern-to-action, reporting chronology, repair closure and incident recovery, and checked where existing oversight might overlap.
The question
How do reported concerns, repair records, notification and recovery connect to evidence of a resolved problem?
What the work revealed
The work separates administrative completion from physical verification and distinguishes reporting from prevention. The archive also defers one broad review offer where an existing review could overlap: the scope needs a genuine uncovered question.
WHAT THIS BRINGS TO A TEAM
Make ownership and closure inspectable while avoiding duplicated effort or an intervention aimed at the wrong constraint.
Evidence, source record & context
Water operations completed records and buyer-research dispositions, September 2026. Historical scope and overlap decisions, not current procurement advice.
Independent research and review proposals. The companies are research subjects, not claimed clients; company indicators do not prove which individual controls failed.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
Turn a promising idea into a testable business case.
My investigation
I investigated the evidence behind the idea and examined whether a published result could transfer to another service. I built zero, partial and full-transfer scenarios, then separated attendance from collected revenue and incremental cost.
The question
Would a reminder intervention create genuinely additional attended appointments and net contribution?
What the work revealed
The resulting model asks what happens when an extra attendance replaces another booking, incurs staff effort or produces no additional collected revenue. A published effect becomes an input to a test, not a guaranteed forecast.
WHAT THIS BRINGS TO A TEAM
Connect research to a decision a business can actually make, with an explicit test before rollout.
Evidence, source record & context
Shapers completed commercial analysis and scenario formula, 18 September 2026. Scenario structure inspected for this portfolio.
The archived external trial result is not a Shapers outcome. Scenario contribution values and transfer assumptions are planning inputs; no revenue improvement or client adoption is claimed.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
I reconstructed the chronology in a published investigation, separated the milestones and compared clock times with narrative durations. I retained discrepancies instead of silently choosing the most convenient version.
The question
Are physical transfer, formal handover and departure being treated as if they were the same event?
What the work revealed
The archived reconstruction distinguishes approximate physical transfer, formal handover and crew departure. One interval is 47 minutes; the recorded total also differs from a narrative duration. Those differences identify questions for the record, not automatic savings.
WHAT THIS BRINGS TO A TEAM
Expose a poorly defined handover so responsibilities, timestamps and the next investigation can be made precise.
Evidence, source record & context
Operational-method completed chronology, 18 September 2026, based on the HSSIB public reference event published in August 2023.
A published reference case, not a case from the four named trust proposals. Approximate times and conflicting statements remain unresolved. Elapsed time is not necessarily avoidable time.
Curated from my retained research records. Original source code, raw datasets and private project documents are not distributed here.
