AI search · agent systems · applied AI · operational evidence
I build systems for the internet that answers back.
I’m Aria Pramesi. I research AI search, build agent-native products, and turn messy operations into systems that can prove what they did. SorcerAI is the public record.
Systems worth explaining.
Selected products and research with explicit status, boundaries, and evidence. The rest of GitHub can remain an unsupervised warehouse.
InboxPolicy
An agent-first email verification and send-decision API that returns an action, confidence, evidence, and cost instead of another vague status field.
Constellation
An evidence-gated AEO, GEO, search, citation, and prompt-intelligence system that preserves provenance and missing evidence instead of compressing everything into a magic score.
Gauntlet
Evidence-first autonomous QA and candidate-bound verification for websites, AI discovery, agent actions, and coding-agent output.
Measured, dated, and public.
AI-crawler findings from the fleet publish as field notes while they mature; interventions become case studies when the before and after are both real. Every number carries a numerator, a denominator, a window, and a source.
Three useful lanes.
I do not sell a mysterious bag labeled “AI transformation.” Engagements start with a bounded problem, evidence, and a measurable decision.
- 01
AI visibility and decision surfaces
Find where a company disappears between search, AI answers, buyer questions, cited sources, and the page that should close the decision.
See the engagement model - 02
Agent-native product systems
Design and build APIs, MCP tools, RAG systems, control planes, and workflows that agents can use without inventing policy as they go.
See the engagement model - 03
Search-to-revenue infrastructure
Turn real demand into useful pages, tools, proof, distribution, and measurable customer actions without feeding the content landfill.
See the engagement model
Short enough to publish.
Build logs, field notes, and occasional rambles. Not every useful thought needs to be stretched into a dead-eyed 1,800-word “ultimate guide.”
AI crawlers spend real budget on dead URLs
Since 2026-07-31, one large content site in the measured fleet has returned 404 to 12,680 of 28,471 GPTBot requests. On this site, 109 of 135 ChatGPT-User requests did the same.
When robots.txt and the edge disagree, the crawler loses
A services site in the measured fleet permits Claude-SearchBot in robots.txt. Its edge rules have since refused 1,999 of that crawler’s requests with 401 or 403.
robots.txt is a request, not a fence
A local-services site disallows ClaudeBot in robots.txt. Since 2026-07-31, 60 of that crawler’s requests were served successfully anyway.
Bring me the expensive mess.
AI visibility, agent infrastructure, or an operation held together by spreadsheets and emotional bargaining.