Safe Journey

Inspiration

Maria is eight weeks pregnant. She does not know it yet. Every morning she applies the same $1.50 hair relaxer she has used for years, bought from the Dollar Tree two blocks from her apartment. The label lists formaldehyde. She cannot read it. No one told her it was a known carcinogen. No one told her it was genotoxic, that it damages the DNA of developing cells, including a baby's.

She is not alone. 94% of severe birth defects occur in low- and middle-income communities (WHO). Low-income maternal occupations are associated with a 51% increased risk of congenital heart defects (PLOS). Babies born to Black mothers are 40% more likely to die from congenital heart defects, with neighborhood poverty cited as a key driver (March of Dimes). Over 40% of women of childbearing age in the US live in fetal care deserts, zip codes where the median income is $53k compared to $76k near specialized care centers (ScienceDirect).

The toxic chemicals in budget products, formaldehyde, phthalates, unrefined petroleum, do not just cause cancer. They are genotoxic. They damage DNA. They cause birth defects. And the people most exposed are the ones who can least afford to know.

We built Safe Journey because the information exists. It is buried in EPA databases, PubChem, and peer-reviewed research. It has never been made accessible to the people who need it most, until now.

What It Does

Safe Journey is a free, AI-powered ingredient safety checker built for pregnant people and low-income communities. You type or photograph a product's ingredient list, a hair relaxer, a lotion, a cleaning spray, and within seconds you get:

A carcinogenicity risk score for each ingredient, powered by a trained XGBoost classifier
A genotoxicity flag identifying chemicals that damage DNA and can disrupt fetal development, powered by a Graph Neural Network (GNN)
A plain-language explanation written at a 6th-grade reading level, no jargon, no paywalls
Multilingual output in Spanish, Korean, Tagalog, Mandarin, and French Creole
Safer alternatives at the same price point, products available at Dollar Tree, Walmart, or Amazon under $5 that do not contain the flagged ingredient
A free clinic locator where you enter your zip code and find your nearest FQHC offering free or sliding-scale cancer and prenatal screenings

On the alternatives: Safe Journey never asks anyone to spend more money. The insight is that safer options already exist at the same stores, on the same shelves, at the same prices. Maria does not need a $30 clean beauty product. She needs to know that the $1.75 version two shelves over does not contain formaldehyde. That is an information problem, not a budget problem. Safe Journey solves the information problem.

How We Built It

The ML Pipeline:

User inputs an ingredient name (e.g. "formaldehyde")
PubChem API returns the SMILES string, the chemical's molecular structure
RDKit computes 130 molecular descriptors from that structure
Our trained XGBoost classifier (trained on 2,020 compounds, F1: 0.76, ROC-AUC: 0.77) outputs a carcinogenicity probability score
A Graph Neural Network reads the molecule as a graph, atoms as nodes, bonds as edges, and outputs a genotoxicity risk flag
Claude API translates everything into plain language, explains the biological mechanism simply, suggests affordable alternatives, and delivers the output in the user's language
HRSA's public FQHC API surfaces nearby free clinics by zip code

Stack: XGBoost · RDKit · PubChem REST API · GNN (PyTorch Geometric) · Claude API · FastAPI · Streamlit

Challenges We Ran Into

Genotoxicity training data is sparse. Unlike carcinogenicity, public genotoxicity datasets are smaller and less standardized. We had to carefully curate and balance the training set to avoid a model that just predicts the majority class.
SMILES lookup failures. Not every ingredient name maps cleanly to a PubChem entry. Trade names, synonyms, and "fragrance" (an intentionally opaque catch-all) required fallback handling and uncertainty flagging.
Plain language is harder than technical language. Getting Claude to explain a genotoxicity mechanism at a true 6th-grade reading level without losing scientific accuracy required careful prompt engineering and iterative testing.
Multilingual output consistency. Ensuring that risk levels and disclaimers translated consistently across five languages without losing nuance took more iteration than expected.
The alternatives problem. Recommending safer products that are genuinely accessible, same store, same price, same product category, required building a curated product database rather than relying on general web results. Generic "clean beauty" recommendations would have defeated the entire purpose of the project.

Accomplishments That We're Proud Of

A working end-to-end pipeline from ingredient name to risk score to plain-language explanation, built in under 90 minutes
A genotoxicity layer that goes beyond existing consumer tools, which focus almost exclusively on carcinogenicity
Genuinely accessible output: multilingual, jargon-free, and actionable within a low-income budget
Alternatives that are actually affordable, not aspirational
A responsible AI framework baked in from day one: confidence scores shown, no data stored, bias in training data disclosed, escalation to human providers always present

What We Learned

The "clean beauty" movement has the right goal and the wrong audience. All the non-toxic product guides in the world are targeted at people who can already afford to make safe choices.
Genotoxicity and carcinogenicity are related but distinct. A chemical can be one without the other. Conflating them in consumer tools creates both false alarms and dangerous blind spots.
The hardest part of health equity work is not the model. It is the interface. The science exists. Making it land for a monolingual Spanish speaker who just wants to know if her shampoo is safe, that is the actual problem.
Recommending alternatives is only meaningful if the alternatives are real. A suggestion that costs more, requires a different store, or is hard to find is not a solution. It is a reminder of what you cannot have.

What's Next for Safe Journey

Expand the GNN genotoxicity model with larger training sets (Ames test databases, ToxCast, DSSTox)
Camera-based OCR so users can photograph a label directly, removing the literacy barrier entirely
Pregnancy-specific risk tier with a separate flag for teratogenic (fetal-harming) compounds, not just carcinogenic or genotoxic ones
Barcode scanning so users scan the product itself, not just the ingredient list
Community health worker toolkit, a version designed for CHWs to use in WIC offices and clinic waiting rooms
Offline mode, because the communities we serve do not always have reliable data connections

The goal is simple. Every pregnant person, regardless of income or language, deserves to know what they are putting on their body.

About Safe Journey

Who We Built This For and Why They Need It

Safe Journey was built for low-income pregnant people, particularly women of color, who rely on dollar stores and budget retailers for everyday personal care products. These communities face a compounding crisis: they are the most exposed to genotoxic and carcinogenic chemicals in cheap products, the least likely to have access to prenatal specialists, and the least served by existing "clean beauty" or "non-toxic" resources, which are designed for and priced toward wealthier consumers.

The data is clear. 94% of severe birth defects occur in low- and middle-income communities. Low-income maternal occupations carry a 51% increased risk of congenital heart defects. Over 40% of US women of childbearing age live in fetal care deserts. These are not abstract statistics. They describe the daily reality of millions of people who have no tool, in their language, at their price point, that tells them whether the product in their hand is safe for the baby they may be carrying.

Safe Journey exists because that tool did not exist before.

How We Used Claude and AI

Claude is the human layer of Safe Journey. The ML models handle the science: XGBoost scores carcinogenicity, the GNN flags genotoxicity. But a probability score means nothing to someone who just wants to know if their shampoo is safe. Claude takes the model output and turns it into something a person can actually use.

Specifically, Claude handles four jobs in the pipeline. First, it writes plain-language explanations of each flagged ingredient at a 6th-grade reading level, explaining the biological mechanism without jargon. Second, it translates all output into the user's preferred language, with Spanish, Korean, Tagalog, Mandarin, and French Creole as priority languages. Third, it recommends safer alternative products available at the same stores and the same price points the user already shops at. Fourth, it provides contextual risk framing, distinguishing between long-term daily skin contact and one-time use, so users can make genuinely informed decisions rather than just feeling scared.

Claude does not replace the science. It makes the science useful.

What Could Go Wrong and How We Addressed It

The model could be wrong. XGBoost and GNN predictions carry uncertainty. Safe Journey shows a confidence score alongside every risk rating and flags low-confidence predictions explicitly. Users are never given a false sense of certainty.

The model could be biased. Our training data skews toward compounds studied in Western research contexts. Chemicals more common in products used by communities of color may be underrepresented. We disclose this limitation directly in the app.

It could be mistaken for a medical diagnosis. Every result includes a clear disclaimer: Safe Journey screens for risk, it does not diagnose disease. Users showing concern about prior exposure are always directed to a human healthcare provider, never kept in the app loop.

Ingredient names could fail to resolve. Not every label ingredient maps cleanly to a PubChem entry. Trade names, synonyms, and vague terms like "fragrance" require fallback handling. When an ingredient cannot be resolved, Safe Journey flags it as unverifiable rather than silently skipping it.

Safer alternatives could be inaccessible. Recommending a product that costs more or requires a different store defeats the entire purpose. Our alternative suggestions are filtered to products under $5, available at Dollar Tree or Walmart, in the same product category.

No data is stored. User inputs are not logged or retained. Safe Journey does not build profiles. It answers the question and forgets.

What We Would Build Next with More Time

The immediate next step is a pregnancy-specific risk tier: a dedicated flag for teratogenic compounds, chemicals with known fetal-harming effects, separate from the carcinogenicity and genotoxicity scores. Pregnant users need a different kind of answer than the general population, and right now the model does not distinguish.

Beyond that, camera-based OCR would remove the literacy barrier entirely. A user should be able to point their phone at a label and get results without typing a single word. Barcode scanning would go one step further, identifying the product before the ingredient list is even visible.

Longer term, a community health worker toolkit would bring Safe Journey into WIC offices and clinic waiting rooms, where it can reach the people who need it most through trusted intermediaries. And an offline mode would ensure the app works in the low-connectivity environments where many of our target users live.

The science is ready. The infrastructure is ready. What Safe Journey needs next is reach.