01 / Software & entrepreneurship
Axyera.
Building trust
into carbon data.
From scattered manufacturing records to evidence-backed product carbon passports.
My role
Solo founder · Product direction · System design
AI-assisted development
Status
Working prototype
Follow the evidence-to-carbon-passport workflow using Axyera's demonstration product.
System overview
One product.
Several connected systems.
- 01
Document intelligence
Multi-format ingestion, AI-assisted extraction, and source-linked review.
- 02
Semantic interpretation
Translating claims into manufacturing activities and relationships.
- 03
Carbon calculation
Supervised factor matching, deterministic computation, and explicit uncertainty.
- 04
Public disclosure
Versioned product carbon passports and supporting methodological context.
01 / THE PROBLEM
Evidence is scattered. The challenge is not.
Product-level carbon accounting can be difficult for manufacturers without specialized sustainability teams. Relevant information is scattered among documents, supplier records, material specifications, energy data, and transport information. Translating that evidence into a meaningful disclosure is not simply a matter of finding numbers.
The challenge I chose to address was making that process more accessible without removing the evidence and uncertainty that give a result its meaning.
02 / What I built
A connected path.
Not a black box.
Evidence → interpretation → a qualified result.
- Authenticated manufacturer workspaces
- Document ingestion and AI-assisted extraction
- Source-linked claims
- Human review and corrections
- Semantic interpretation of manufacturing activities
- Supervised emissions-factor selection
- Deterministic carbon calculation
- Public, versioned carbon-passport disclosures
Manufacturer Documents
Extract & Structure
Human Review
Factor Resolution
Carbon Calculation
Public Passport
03 / CONTEXT MATTERS
When the number is right, but the meaning is wrong.
Extracting a number from a manufacturing document does not establish what that number represents. The capacity of a delivery vehicle is not the mass of a particular shipment. A supplier's location is not necessarily the manufacturer's production location. A finished product's weight is not interchangeable with every material quantity mentioned in its supporting records.
As I identified these failure cases during development, I directed changes in how Axyera interpreted document-backed information. The system evolved beyond treating every extracted quantity as an immediate calculation input, introducing a layer that distinguishes manufacturing activities and their relationships.
The distinction I found most important was simple: accurate extraction is not the same as accurate interpretation.
- 01Source-linked evidence each claim points back to the document it came from.
- 02Activity interpretation extracted values are read as manufacturing activities, not raw numbers.
- 03Claim status what has been accepted, corrected, or is still pending.
- 04Human review a person confirms the interpretation before it can inform a calculation.
04 / FACTOR SUITABILITY
The closest match is not always the right one.
Even a correctly interpreted manufacturing activity can lead to a misleading result when paired with an unsuitable emissions factor. Two reference records might describe superficially similar materials but differ in geography, production process, physical form, or accounting boundary.
I wanted Axyera to account for those distinctions instead of treating factor selection as a simple lookup. Its current resolution system evaluates suitability across several dimensions and records human decisions about proxies, alternatives, and exclusions.
Sometimes a proxy is useful for an explicitly qualified screening estimate. Sometimes the evidence does not justify using it at all.
- 01Factor provenance the reference record behind each decision.
- 02Proxy decisions recorded, and qualified when used.
- 03Alternatives and exclusions candidates kept for reference only.
05 / UNCERTAINTY
Knowing when not to give a number.
The apparent simplicity of a single carbon-footprint figure conceals a difficult question: what happens when part of the underlying evidence is incomplete?
Axyera separates AI-assisted extraction from deterministic calculation. The engine accounts for identified unresolved activities, distinguishes evidence traceability from calculation readiness, and can withhold an aggregate result when the identified components do not support it.
This does not guarantee a complete inventory of every possible manufacturing activity. But it reflects an important product decision: a system should not make uncertainty disappear merely because a neat result would be easier to display.
- 01Traceability vs. readiness evidence can be traceable yet not ready to calculate.
- 02Unresolved inputs identified gaps stay visible instead of disappearing.
- 03Activity breakdown the total decomposed into its contributing activities.
- 04Factor provenance the source of every emissions factor applied.
06 / PUBLIC CARBON PASSPORTS
From private evidence to public understanding.
Axyera brings the reviewed information together in a product-level disclosure intended to be accessible through a public link or QR code. Passports communicate the calculation, supporting evidence categories, factor provenance, and relevant qualifications.
Publication can create a new versioned record as product information changes. The platform distinguishes manufacturer-declared and evidence-screened information from independent external verification.
07 / Building through iteration
A product shaped by the problems it revealed.
What began as a way to make product carbon accounting more accessible became a much more demanding engineering challenge: translating fragmented, imperfect manufacturing evidence into calculations that could be examined and justified.
As I developed Axyera, I repeatedly encountered cases where a technically valid result concealed a conceptual error. A number extracted correctly could represent the wrong activity. An emissions factor could appear relevant while describing a different production process. A complete-looking footprint could obscure unresolved assumptions.
Each of these discrepancies changed not simply what Axyera displayed, but what its results could responsibly claim.
Development approachFounder-directed product architecture and iteration, supported by AI-assisted coding tools, GitHub version control, and targeted regression checks.
Axyera / Behind the build
The decisions behind the system.
Building Axyera taught me that producing an answer and being able to justify it are two different engineering problems. Three decisions fundamentally changed how I approached the platform.
- 01
When accurate extraction produces the wrong answer
Early in development, I encountered a problem that couldn't be solved simply by improving document extraction: the system could identify a number correctly while misunderstanding what it represented.
Rather than treating extracted values as calculation-ready data, I directed the introduction of a semantic interpretation layer to distinguish manufacturing activities, quantities, and their relationships.
What I learnedTechnical accuracy is only meaningful when the underlying interpretation is correct.
- 02
Why the closest match wasn't always the right one
Emissions factors introduced a different challenge. A reference dataset could contain a seemingly suitable match that actually described another material, manufacturing process, or geographic context.
I moved away from treating factor selection as an automatic lookup and toward supervised resolution, where suitability, provenance, and the justification for using a proxy could be examined before calculation.
What I learnedA confident-looking result can conceal assumptions that deserve more scrutiny, not less.
- 03
Making uncertainty part of the product
I realized that a product could have well-organized supporting documents without having sufficient evidence for every component of its carbon footprint. Evidence traceability and calculation readiness weren't the same thing.
This distinction shaped the platform's approach to unresolved inputs, review states, and public disclosures, so that a carbon passport could communicate the limits of its underlying evidence rather than obscure them.
What I learnedTrustworthy software must communicate what it doesn't know as carefully as what it does.
- 01
From documents to reviewable evidence
Connected multi-format document processing with source-linked claims and human review, making extracted information inspectable before calculation.
- 02
From extracted values to interpreted activities
Directed the development of semantic interpretation and supervised factor resolution to address ambiguous quantities, supplier relationships, and unsuitable reference data.
- 03
From prototype iterations to a structured development workflow
Established a GitHub-based development process with targeted regression checks, continuous integration, and controlled deployment.
08 / Results & next steps
Built to question the answer.
Axyera does more than produce a carbon-footprint figure. It seeks to make the evidence, assumptions, and uncertainty behind that figure intelligible—so a result can be examined, not simply accepted.
Entrepreneurial recognition
Youth Business Venture Competition
Global Top 15
Global Top 15 finalist. Only Romanian finalist; presented Axyera at the competition final held at Stanford University.
GENIUS Olympiad
Entrepreneurship finalist
Entrepreneurship finalist. The only female finalist representing Romania.
Next priorities
- Manufacturer testing and feedback
- More reliable handling of ambiguous evidence
- Improved calculation and publication safeguards
- Development across additional product categories
Axyera is a working prototype, with manufacturer testing and broader product coverage as the next stage of development.