Personal details

Roman P. - Remote data scientist

Roman P.

Based in: 🇺🇸 United States
Timezone: Central Time (US & Canada) (UTC-6)

About

Engineer building production AI systems and on-chain software, with a specialism most builders skip: proving the
thing does not break. I write invariant and fuzz harnesses — properties that must hold no matter what sequence
a stranger calls — and carry that discipline from smart contracts, where a mistake is priced in dollars immediately,
into agent and API work.
Unit tests call functions in the order the author thought of. Production runs in the order an attacker thought of.
Everything below is verifiable: open repositories, run counts, accepted audit findings, public certificate links.

Work Experience

Founder and principal engineer
AutoMate AI — Elite Product LLC | Nov 2024 - Present
Ethereum
Solidity
Threat Modeling
API Integration
System security
LLM
Smart contract
Software agents
Ai system engineer

Delivery of agent, integration and audit work for business clients; written specifications in, working builds and written daily updates out.Founder and principal engineer. Elite Product LLC registered in California on 9 November 2024 (file no. 202464515740).

Projects

Open Counter — a sales agent its owner can teach
2026 | https://github.com/chainwright/open-counter
Python
PayPal
TypeScript
React
LLM
A shop for a one-person maker with an AI sales agent that sells over chat and over an MCP server, with multi-item PayPal checkout and made-to-order invoices. The owner corrects the agent in plain words (or a second model drafts the lesson from chats where it got stuck), and every later chat is checked in code for whether the agent followed it, with the order value of those chats. Live demo: open-counter.onrender.com. 123 tests.
Agent Grader — evaluation service for agent output
2026 | https://github.com/chainwright/agent-grader
Python
PostgreSQL
Docker
Fastapi
LLM
Scores agent output against declarative rubrics. Rubrics are validated when a task is created, so a bad criterion cannot pass for an agent failure; every criterion gets its own verdict with a failure breakdown; retries are deduplicated by idempotency key so a timeout cannot inflate the pass rate. Async SQLAlchemy, 34 tests in CI against a real database.