Back to the shelf

033

MicroStats

An audit-ready QC web app for two potency assays, where every result is calculated twice and both must agree.

Launched · Started March 2026 · System · Vibe-Coded

That's the whole thing.

The site's sound is off. The narration still plays.

This reading didn't load. The transcript is below.

Read the transcript

A validated Excel workbook sat at the center of a quality control lab. It worked. But for high-stakes calculations, a spreadsheet is not a system. That tension is what started MicroStats.

MicroStats evaluates two fixed potency assays. Every result is calculated twice, independently, by the system itself. Both calculations must agree before anything moves forward. The math runs on log-transformed linear regression, with hard suitability checks and a readable trace of every step.

Access is role-based. The audit trail is immutable. A portal of validation documents sits alongside the app.

I started this in March twenty twenty-six, for the pharma company I work with. Vibe-coded with AI. It has launched. Follow along if that interests you.

Narrated in my own cloned voice, generated with ElevenLabs v4.

A validated Excel workbook sat at the center of a quality control lab. It worked. But a spreadsheet doesn’t enforce and it doesn’t trace, and when a calculation carries real consequences, that distinction matters.

MicroStats replaces that workbook. It’s an internal web app for the lab at the pharma company I work with, built to evaluate two fixed potency assays with the guardrails a regulated environment demands.

Guided entry walks you through each assay. Log-transformed linear regression runs beneath. Hard system-suitability checks stop the calculation if the data doesn’t qualify. Every result is calculated twice, independently, and the two must agree. That’s the shadow verification.

The output is readable. A calculation trace you can follow step by step, a residual plot, HTML and PDF reports. Each assay moves through a lifecycle from draft to submitted to final, with mandatory comments. Return and superseding-version workflows are built in.

Access is role-based: doer, approver, administrator. The audit trail is immutable. Tamper evidence is part of the record. A help manual lives inside the app. A portal of validation documents lives there too.

I started this in March 2026, vibe-coded with AI. It’s launched.

The making

  • Node.js
  • SQLite

01 · The problem

Before it existed

What hurt

A validated Excel workbook carried the potency statistics for two assays. It gave right answers, but kept no audit trail and sat one paste-error from breaking.

Who uses it

  • QC analysts
  • QC approvers
  • Administrators
  • 76commits
  • 2fixed assays
  • 0.98R² gate, a hard stop
  • 1e-9two calculations must agree to

02 · The flow

6 steps, in the order they happen

  1. 01 / 06

    Sign in by role

    You
    Logs in as analyst, approver or administrator
    It answers
    Shows only what that role may do
    The MicroStats sign-in card with username and password fields
  2. 02 / 06

    Enter the assay

    You
    Picks the assay and types the observed readings
    It answers
    Calculates means and log concentrations as each row is filled
    Guided entry form with assay metadata and a table of standard responses
  3. 03 / 06

    Run the verified calculation

    You
    Presses one button
    It answers
    Computes every result twice and refuses to go on if they disagree
    Result review with slope, intercept, R squared and the final activity
  4. 04 / 06

    Inspect the curve

    You
    Looks at the fit and the residuals
    It answers
    Draws the standard curve, the test points and the error summary
    A standard curve with reference points, test points and a fitted line
  5. 05 / 06

    Review and sign off

    You
    Approver opens the pending record
    It answers
    Shows the audit trail; approve, return with a reason, or supersede
    Approver view with a pending record and its audit trail of events
  6. 06 / 06

    Manage the accounts

    You
    Administrator opens the admin portal
    It answers
    Creates users, resets passwords and logs every session
    The admin portal listing three role accounts and their status

03 · What was special

What it stands on

  1. A hard stop, not a warning

    Below an R² of 0.98, the system refuses to proceed. It’s not a warning; it’s a stop.

  2. Shadow verification

    A second, independent calculation path must match the first to 1e-9, or the result does not exist.

  3. Validation inside the product

    The user manual and the validation documents sit inside the app, open to read without a login.

Every result is calculated twice, and both calculations must agree.

Making-Of

The spreadsheet worked fine. That's exactly why I replaced it.

Recorded voicenote – the decision to replace CombiStats with self-built, GMP-validated software (private archive)

  1. The world before

    The QC microbiology lab at Venus Remedies ran anticoagulant potency statistics – Anti-factor Xa and Anti-factor IIa – on two systems: CombiStats, EDQM's licensed statistical software, and an internally validated Excel workbook. The workbook produced correct answers and lived in the zone every GMP practitioner recognizes: validated on paper, fragile in practice. Manual cell references. No audit trail. One silent paste-error away from broken.

  2. One day

    Mid-March 2026. A voicenote, recorded before my rational brain could talk me out of it: convert that Excel into real, GMP-validated software. Same purpose and workflow as CombiStats. Validated against the Excel itself. Auditable by design. The instruction I gave myself: act and think like a GMP auditor. I had never built GMP-compliant software before.

  3. Because of that

    1. 13 March 2026. A two-to-three-day build. The statistics re-derived from first principles – LOG10 linear regression, geometric-mean potency, R squared as the system-suitability gate – because 'I think I remember this' is not a GMP defense. Below 0.98, the system refuses to proceed. Not a warning; a stop.
    2. The regulatory framework learned from scratch: GAMP 5, EU GMP Annex 11, the US FDA's electronic-records rule – each read twice, once as a developer and once as an imaginary auditor.
    3. Claude Code as pair-programmer throughout. The design, workflow logic, validation decisions, and statistical approach were mine; Claude Code wrote the code I directed. Node.js, SQLite, hosted on Railway. The architecture is not complicated. The discipline behind it is.
    4. Shadow verification: every result computed by a second, entirely independent calculation pathway, the two agreeing to within 1e-9. If they disagree, the result does not exist. Draft, Submitted, Final lifecycle, Doer / Approver / Administrator roles held separate, immutable audit trail with SHA-256 tamper evidence, admin portal logging every session and update.
    5. The validation documentation written for this software, this lab, this workflow – URS, FS, DS, IQ, then OQ and PQ – plus the user manual. Not delegated. Not adapted from a template.
    6. A first demo at home landed flat; the math was questioned. That sting became fuel – by the time the team's calculation owner ran his independent verification on the live system, there was no gap to find.
  4. Until finally

    The system went live in the GMP environment and the microbiology team entered more than 15 real potency assays. Prince, the calculation owner, independently verified the math – identically. MicroStats was formally GMP-implemented in QC Micro. Then the designer's pass: the generated report redesigned with a Venus-emblem banner and signatures rendered in Dancing Script – blue for the lab analyst, black for the QC approver – so a digital report reads as signed, not stamped. It now runs alongside CombiStats for this workflow at a saving of over 600 euros per year.

  5. Since then

    MicroStats became the proof-of-pattern – Granth, Dwaar and others followed, a wave of disciplined, personalized software at Venus. The validated-calculation approach was later generalized into the quality system's LIMS calculation engine, and the company's technical team adopted GMP-oriented software development into its 2026 mission. The software is still maintained: a recent pull request corrected a single validation-document reference (PQ-07, Format No. F/QC/546/01-00) with exactly the precision the original build demanded. The maintenance record is part of the proof.

The evidence Process, hands, and artifacts

Process

  1. Origin

    Replace the CombiStats license and the validated Excel workbook with self-built GMP-validated software – act and think like a GMP auditor

  2. Build

    Two-to-three-day build – LOG10 regression, R-squared gate at 0.98, shadow verification to 1e-9, RBAC, SHA-256 audit trail; Node.js + SQLite on Railway

  3. Validation

    Full validation documentation authored – URS, FS, DS, IQ, then OQ/PQ – plus the user manual

  4. Go-live

    Live in the GMP environment; 15+ assays on day one; independent math verification by the team's calculation owner

  5. Design pass

    Report redesigned – Venus-emblem banner, Dancing Script signatures (blue analyst, black approver)

  6. Maintenance

    PQ-07 validation-document reference corrected – Excel Sheet QC-CAL-033-00, Format No. F/QC/546/01-00

Hands

Akshansh

  • Concept and direction
  • Statistics review and derivation
  • GMP workflow design
  • Validation documentation (URS through PQ)
  • Report design
  • Every product decision

Agents

  • Claude CodePair-programming execution

In a regulated environment, the process is the record, the record is the evidence, and the evidence is the only thing that separates authorship from assertion.