R / R RESEARCH PORTFOLIO
ANTHROPIC RESEARCH RESIDENT · PORTFOLIO

Building systems that can be checked

I am an entrepreneur and project initiator interested in how people can verify the work of AI agents against evidence, not just their reports.

From business processes to agent workflows

I have been an entrepreneur since 2003. I founded an advertising agency that operated from 2003 to 2010 and participated in developing a car service business from 2004 to 2010. In 2007, buildings for the service and an auto-parts shop were constructed to my project brief, and the physical shop opened.

01

Port3 2010–2018

I financed the project and defined its order, supply, payment, document and deferred-payment control processes, along with customer interfaces. Programmers implemented the software. The website and software system were sold in 2018.

Our team had 15 people; monthly reach exceeded 1.5 million unique visitors, with over 77,000 registered customers and more than 1,000 active suppliers. Peak database traffic exceeded 650 accesses per second, including accesses caused by bots and search engines. These are database accesses, not users, orders or HTTP requests; suppliers are not counts of individual API integrations.

02

R2R.os IN DEVELOPMENT

I initiated R2R.os as a practical environment for studying agent behaviour, context retention, delegation, verification and delivery. I define goals, rules and acceptance criteria and examine failures. Full autonomy and a measured reliability improvement have not been established.

03

RealtyAI DOCUMENT PROTOTYPE

I initiated RealtyAI. Document preparation is being prototyped. A customer in Russia is ready to test the prototype with employees. Property and counterparty checks, calls, dialogue, publication materials, examinations and further staff training are intended functions, not confirmed implementations. Pilot readiness is not a completed successful pilot.

04

AgentChat × BezProstoya AGENTCHAT: NEAR COMPLETION · BEZPROSTOYA: PILOT READY

AgentChat is being built as the system’s “hands”, intended to act on a customer’s computer. Under R2R.os it is intended to serve as the remote agent for BezProstoya. BezProstoya has a customer and is ready for pilot testing; a successful pilot, autonomy and uninterrupted operation are not established.

A real run, fictional documents

The downloadable package contains entirely synthetic scans, a blank template, expected values and a DOCX produced by a real prototype run on 14 November 2026 at 23:10, version 493624c. The run received final acceptance: the surname was correct, an ambiguous answer produced a warning, and confirmation led to the final DOCX. Check the result against the sources; this example does not establish general reliability.

Where verification matters

Component progress did not equal delivery

Task. Prepare and deliver a contract from a fictional document set. Observed discrepancy. In a 13 November 2026 live test, documents were received and a partial draft appeared, but template parsing repeatedly failed and no final contract was produced in that run. Verification requirement. Open and inspect the final file obtained through the actual user path, including its fields and source references. Not yet proven. A later successful reduced-set run does not prove general reliability or completion for every input.

Acceptance cases lost in a handover

Task. Fix answer parsing while preserving previously accepted cases. Observed discrepancy. At handover to a new task, protected cases from the earlier review were absent from its acceptance criteria. Four successive rounds stopped on contradictions with those cases. The manager issued three addenda; the rule was made additive, after which the result was accepted. This is the observed sequence, not a proven causal diagnosis. Verification requirement. Carry protected cases into the acceptance criteria and check the new rule against them, including existing warnings and literal answers. Not yet proven. Carrying cases forward prevents recurrence in general.

What does verification actually catch?

I propose comparing three review conditions: report alone; report plus log; and independent checking against source documents. Measure missed errors, false rejections, correctly completed tasks, cost and human time. Customers ready to pilot RealtyAI and BezProstoya can help assess scenario realism; pilot readiness is not an experiment result.

Read the bilingual proposal →
AI assisted editing and translation. · English and Russian versions describe the same facts.