Humanware← madhukamath.com
Case study · Health-record intelligence

Turning a shoebox of scattered medical files into an auditable, private health record.

A participant-owned system that indexes years of inconsistently-named scans, pathology and treatment documents into a searchable, source-traceable archive — while preserving uncertainty and keeping personal health information private by design.

Open the live demo → See how it works Private access — by request
Data engineering Information architecture Privacy by design Auditability
The problem

The information existed. The record didn't.

A decade of medical files had accumulated across desktops, downloads, cloud folders and email exports — as photos, scans, PDFs and screenshots with names like Current CT CAP 2.jpeg and Report 1.png. The files were present, but no one could answer the questions that actually matter: what happened, in what order, from which document, and what still needs attention. The burden of reconstructing the story fell entirely on the patient.

10 yrs
of records spread across many folders
Mixed
photos, PDFs, screenshots, exports
No index
no dates, no sequence, no provenance
Before / after

From raw files to a structured index

The first phase is deliberately organizational, not clinical. It builds a trustworthy metadata layer before anyone attempts to interpret a single medical finding.

⛰ Before — a file pile

  • Ambiguous names, no consistent dates
  • Pages, parts and second opinions loose
  • Requests indistinguishable from results
  • Duplicates and near-duplicates unknown
  • No provenance, no way to search

✓ After — an auditable index

  • Honest dates with explicit precision
  • Pages grouped into document families
  • Requests separated from reports
  • Byte-identical duplicates detected (SHA-256)
  • Every row searchable and source-linked

All figures and screenshots on this page come from a synthetic demonstration dataset — invented patient, invented providers, sample filenames. No real person's health information appears anywhere here.

Demonstration figures (synthetic)

What the index exposes

Running the pipeline over the synthetic sample of 34 files produces a manifest-backed dashboard. Crucially, it surfaces its own incompleteness instead of hiding it.

34
source files indexed
25
provisional document families
2
byte-identical duplicates (1 group)
16 / 34
dated — only 6 to an exact day
26%
provider coverage (honestly shown)
0
clinical fields interpreted (by design)
Key design decisions

Judgement, not just parsing

The interesting work was refusing to let a polished dashboard imply more certainty than the data supports.

1

Count files, not "records"

A scan and its second-opinion copy aren't two clinical events. The headline counts source files, with separate figures for unique content and document families.

2

Honest dates over false precision

A filename that says only "Jun 2019" is shown as Jun 2019 with a month precision tag — never fabricated to the 1st of the month, which could reorder a treatment timeline.

3

Requests ≠ results

A test request is not proof a test happened. A status field separates requests, reports, discharge summaries and second opinions.

4

Exact duplicates vs. document families

SHA-256 flags only byte-identical files. Everything else — re-scans, re-exports, pages — is grouped as a provisional family with a stated basis and confidence, pending review.

5

Show the gaps

Undated files and missing years are surfaced as "completeness unknown," not quietly dropped or shown as a false timeline event.

6

No clinical interpretation without verification

Phase one reads filenames and hashes only — never document contents. Confidence is capped until a separate, explicit review phase.

Auditability

Every grouping is explained

Document families are never asserted silently. Each shows how it was formed and how much to trust it — reconcilable against an exported manifest (synthetic sample shown).

FamilyFilesBasisConfidence
CT series4Explicit sequence numbering in filenamesMedium — needs content check
Histopathology report2"Report 1 / 2" numberingMedium — needs content check
Radiation discharge summary2"Summary 1 / 2", same monthMedium — needs content check
Blood results (copy A / B)2Identical SHA-256 hashHigh — byte-identical duplicate
Current CT CAP2Similar filename onlyLow — needs content check
In the demo

Searchable, filterable, sortable

🔎

Full-text search

Filter by filename, provider, modality or body site — instantly, across the whole index.

🎛

Combinable filters

Chips for Dated / Undated / Requests / Duplicates stack with search to narrow to exactly what matters.

Sortable columns

Sort by date, status, domain, modality or provider — month-only and undated records order predictably.

🧩

Multi-modality aware

A single "Mammogram, Ultrasound, Biopsy" file surfaces all three modalities, not just the first.

Try the interactive demo →
Privacy architecture

Three audiences, three access tiers

The same work serves care and demonstrates capability — without making a private health story the price of being seen. Access is layered, not all-or-nothing.

Doctors Private
Real record + source documents
  • Approved-email one-time codes
  • Per-clinician, individually revocable
  • Documents served only behind auth
  • Access logged; no shared password
Collaborators Redacted
Working demo, no PHI
  • Invented names & providers
  • Sample filenames only
  • Full features, inert file links
  • Isolated from real data
Everyone Public
This case study
  • Problem, approach, results
  • Design & privacy decisions
  • No names, records or health details
  • Automated leak-gate on build
Why it matters professionally

What this demonstrates

🧭

Systems thinking under ambiguity

Designing a trustworthy structure from messy, real-world inputs — and knowing what not to claim.

🔐

Privacy-first design

Access tiers, data minimisation and provenance treated as architecture, not an afterthought.

📐

Auditability & rigour

Every displayed number reconciles to an exported manifest; every grouping states its basis and confidence.

🗣

Lived-experience insight

Built from the patient's side of the system — the perspective institutions rarely design for.

See it working

Explore the searchable, source-traceable index built on a fully synthetic dataset.

Open the demo →