One of my recent side projects is Before You Trust, a public-web research tool designed around a simple idea: before drawing conclusions about a person or company, make sure you are looking at the right identity and the underlying sources.
The interesting technical decision is that the core search workflow does not depend on an LLM. Instead, it focuses on search coverage, identity matching, evidence filtering, and explicit user confirmation.
That constraint made the project a useful exercise in search engineering.
The problem with name-based web search
Searching for a person by name is easy. Determining whether every result belongs to the same person is not.
Common names create obvious collisions, but even distinctive names can return:
- social profiles for different people;
- copied or syndicated articles;
- outdated pages;
- low-quality directory listings;
- results that mention the name without identifying the person;
- search snippets that look relevant but point to weak evidence.
A system that immediately summarizes all of those results risks producing a polished answer about the wrong person.
So I designed the workflow around identity first, research second.
Step 1: search multiple public-web sources
Before You Trust uses self-hosted search infrastructure based on SearXNG and YaCy. The purpose is not to maximize the number of URLs. It is to obtain enough independent coverage to build plausible candidate identities without relying on a paid per-search API.
Results are normalized, deduplicated, bounded, and filtered before they reach the next stage.
This is similar to data engineering in miniature: raw search results are inputs, but they are not yet trustworthy data.
Step 2: separate likely identities
The application groups evidence into candidate identities and asks the user to explicitly choose “This is them” before deeper research continues.
That interaction is important.
The system does not silently assume that the top search result is correct. The user remains part of the identity-resolution step, and deeper searches are anchored to the selected candidate.
For an evidence-heavy tool, this is more useful than pretending identity matching is perfectly automatic.
Step 3: filter aggressively
Search systems often look better when they return more results, but quantity is not the same as quality.
For this project I deliberately prefer conservative filtering. Low-confidence results that may belong to a namesake should not be promoted into the final report simply because they match a keyword.
The final output is therefore closer to a sourced research brief than a generic search result page.
Step 4: preserve evidence instead of generating a score
I also avoided a single “trust score” or “danger score”.
A numeric score can create false precision. Public-web evidence is incomplete, context dependent, and sometimes wrong. A result can be relevant without proving a claim, and an absence of results does not prove that somebody is safe.
The useful unit is the source itself: what was found, where it came from, and why it appears connected to the selected identity.
Privacy was an architectural requirement
Ordinary searches are transient rather than stored as a permanent search-history database. Search and report request bodies are excluded from application logs and analytics, and optional report delivery is only triggered when a user explicitly requests it.
The project also includes consent-aware analytics and clear limitations around appropriate use.
These choices add implementation work, but they are part of the product rather than legal text added at the end.
What this project taught me
Before You Trust reinforced a lesson I have seen in document processing and data pipelines as well: reliability often comes from controlling the boundaries around a model or search engine, not from making the central algorithm more complicated.
The important pieces are:
- normalize the input;
- keep provenance;
- reject uncertain matches;
- make assumptions visible;
- let a user verify ambiguous identity decisions;
- test the full workflow, not only one function.
Those principles also apply to RAG systems, OCR pipelines, entity resolution, and structured extraction.
Before You Trust is still a public-web research assistant rather than a background-check service, and its coverage can be incomplete. But as an engineering project, it has been a useful way to explore how far careful search, filtering, and product design can go without using AI as a shortcut for every step.