Problem

• Search returned companies, but the ranking didn't answer the client's question
• Fit and data quality were blended into one confusing score
• Companies with detailed websites ranked high even when they were a poor fit
• Asking for a set number of companies often returned far fewer useful ones
• Researchers couldn't easily see why a company was included
• The team was spending more time double-checking results than using them.

Solution Exploration

• Keeping a single blended score and fine-tuning it
• Letting AI decide the filters and scores on its own
• Splitting the work: AI does the research, while clear, fixed rules handle filtering, fact-checking and scoring

How Dezoko Helped

• Two separate scores: Query Match shows how much of the brief was confirmed, and Data Confidence shows how well supported each company's information is.
• Smarter ranking: Results are ordered by match level (Exact, Probable, Related) first, then Query Match, then Data Confidence.
• The right number of results:
  • The search gathers more candidates than requested.
  • The limit is applied only after checking and ranking.
  • Checking stops as soon as enough qualifying companies are found, so no research effort is wasted.
• Evidence for every fact: Details such as employee counts come with a supporting quote, and every fact is checked against the website it came from.
• Honest about gaps: Information a company hasn't published is marked as unverified, while a company that clearly fails a must-have requirement is removed.

Transparency and Control

• Why each company matched the brief
• Which evidence supports each fact
• Which facts are missing or unverified
• Separate scores for fit and data quality
• Instant exports from saved results, with no need to re-run the research

Key Outcomes

• Search results now show match quality and data confidence separately
• Researchers can see the reasons and evidence behind every match
• Missing facts are visible instead of hidden
• Requests return the number of qualifying companies the team asked for
• Exports are instant because they reuse saved results instead of re-running the AI

How We Built It

We rebuilt the ARGOS search as a clear five-step process, with AI doing the research and fixed rules making the final calls:

A Clear Five-Step Search

• Understand the brief: AI reads the research brief and works out exactly what the client is looking for.
• Find and research: AI searches the web for candidate companies and pulls key facts from their websites.
• Filter out poor fits: Must-have requirements are applied by fixed rules, not AI judgement.
• Check the evidence: Every fact is checked against the website it came from.
• Rank and save: Results are ranked and stored, ready to review or export.

Every Fact Checked Against Its Source

• AI handles what it does best: understanding the brief, finding companies and reading their websites.
• Filtering, scoring and fact-checking follow fixed, predictable rules.
• If a fact can't be confirmed on the company's own website, it is clearly marked as unverified.

Built to Change Safely

• The website, the core system and the AI researcher are separate parts, each of which can be updated on its own.
• Companies, datasets and search jobs are all kept in one secure database.

What We Learned Along the Way

• We started with a single combined score:
  • Mixing fit with data quality let well-documented but irrelevant companies rise to the top.
  • We split it into two clear scores, Query Match and Data Confidence.
• We first fetched only the number of companies requested:
  • When some failed the filters, too few useful results were left.
  • The search now gathers extra candidates and applies the limit after checking and ranking.

Built With

• AI: Google Gemini
• AI framework: Google Agent Development Kit
• Web app: Next.js, FastAPI
• Database: PostgreSQL
• Hosting: Google Cloud

Business Impact

• Rankings reflect how well a company fits the brief, not how much it publishes about itself
• Every result can be traced back to the evidence behind it
• Checking candidates in batches avoids unnecessary AI work
• Each part of the system can be improved without disrupting the rest