About the role

About the role
BluPax is a pharmaceutical distributor. A lot of what keeps the business running is still done by hand — people pulling data out of one system every morning, checking it against another, and re-entering the result somewhere else. We are building a development team to change that, and this is the role that does it. You would build the automation layer: agents that run on a schedule, connect to the systems we already use, do the work, and escalate only the exceptions that need a person. Not a chatbot. Not a pilot. Software that runs unattended and that people come to rely on.
The first project
Our compliance team handles controlled substances, and two of their daily processes are the first targets. One is a high-volume inventory report assembled by hand every morning before the day starts. The other is an order review that checks proposed orders against customer and sales history under federal thresholds

What you would do

• Build scheduled, autonomous pipelines: fetch the data, do the work, produce the output, deliver it — without a human starting it
• Design the exception path. Clean cases should complete silently; the ones that need attention should reach the right person with the context to act
• Connect to the systems we run — our ERP, our data platform, third-party tools, and web interfaces where no API exists
• Build the deterministic guardrails and audit trail underneath all of it — input validation, output checks the model cannot influence, and traceability from any result back to the records and logic version that produced it
• Capture user corrections as a feedback signal, so the system’s accuracy is measurable and improves
• Work directly with the people whose jobs you are changing, and understand a process properly before automating it
• Leave behind something maintainable — documented, in source control, and operable by someone who is not you

What we need

Real experience building agent systems that run unattended
This is the requirement we cannot flex on. We are not looking for someone who has used AI tools, or built a chat interface over a document set. We are looking for someone who has built harnesses — a model with tools, memory and a task loop, running on a schedule, doing work without a person in the conversation. Tell us about one you built and shipped. What it did, what it got wrong, how you found out, and what you changed. We are as interested in the failure handling as in the capability.
The discipline that makes a model safe to leave running
Some of what you build here touches federally regulated processes. Getting an answer mostly right is not good enough, and neither is escalating everything and calling it safe.
What separates an automation we can trust from a demo is the scaffolding around the model, and we have strong views on it:

• Deterministic checks bracket every model step. Validate the inputs before the model sees them, and validate the output against rules that do not involve a model. The reasoning sits in the middle, and it never certifies its own work — a verdict on correctness comes from a check the model cannot influence.
• Missing is a failure, not a pass. An empty result, a partial record, a field that did not populate — these fail loudly. The most expensive bugs in this class of system are the ones where nothing appeared to go wrong.
• Every output traces to the exact inputs that produced it. When a number is questioned six weeks later, you can show which records, which version of the logic, and which run produced it.
• Confidence thresholds are explicit. What the system handles alone, what it flags for a person, and what it refuses to decide — decided deliberately and written down, not emergent.
  You do not need to have used those words. You do need to have learned the lesson some other way, and to recognize why it matters more here than in most places. If your instinct on finding compliance logic in an unversioned script is that it needs fixing before it needs extending, you will do well here.
  Our stack — ideally you know some of this
• Python or Node for the pipelines. We are not precious about which
• Agent harnesses — the orchestration layer around the model, not just the model. Claude Code and similar coding agents, especially used across a team; custom harnesses you have built yourself; anything where you have had to decide what the loop does, where it stops, and what happens when a step fails. This is the part of the stack we care most about
• Agent frameworks and tool protocols — MCP, function calling, whatever you have used to give a model real capability
• Browser automation — some of what we need to reach has no API
• SQL and analytical databases — our data platform is ClickHouse
• ERP data — SAP Business One, though you do not need to have worked in it
• Source control and CI as habits, not as process imposed on you

Nice to have

• Pharmaceutical, medical or another regulated distribution environment
• Experience automating finance operations — AP matching, payment processing, reconciliation
• Document and data extraction from inconsistent third-party formats
• Having built something that a non-technical team then used every day without you

The engagement

• Six months, with a review at month four and the option to extend
• Full-time for the six months, as an independent contractor
• $50–75 per hour (USD), depending on depth of relevant experience
• Fully remote from Latin America, with at least four hours of daily overlap with US Eastern time. Our compliance and operations people are in New Jersey and the processes you are automating happen in their mornings
• Professional working English. You will be interviewing the people whose work you are changing, not just reading specifications

About the role

About the role
BluPax is a pharmaceutical distributor. A lot of what keeps the business running is still done by hand — people pulling data out of one system every morning, checking it against another, and re-entering the result somewhere else. We are building a development team to change that, and this is the role that does it. You would build the automation layer: agents that run on a schedule, connect to the systems we already use, do the work, and escalate only the exceptions that need a person. Not a chatbot. Not a pilot. Software that runs unattended and that people come to rely on.
The first project
Our compliance team handles controlled substances, and two of their daily processes are the first targets. One is a high-volume inventory report assembled by hand every morning before the day starts. The other is an order review that checks proposed orders against customer and sales history under federal thresholds

What you would do

• Build scheduled, autonomous pipelines: fetch the data, do the work, produce the output, deliver it — without a human starting it
• Design the exception path. Clean cases should complete silently; the ones that need attention should reach the right person with the context to act
• Connect to the systems we run — our ERP, our data platform, third-party tools, and web interfaces where no API exists
• Build the deterministic guardrails and audit trail underneath all of it — input validation, output checks the model cannot influence, and traceability from any result back to the records and logic version that produced it
• Capture user corrections as a feedback signal, so the system’s accuracy is measurable and improves
• Work directly with the people whose jobs you are changing, and understand a process properly before automating it
• Leave behind something maintainable — documented, in source control, and operable by someone who is not you

What we need

Real experience building agent systems that run unattended
This is the requirement we cannot flex on. We are not looking for someone who has used AI tools, or built a chat interface over a document set. We are looking for someone who has built harnesses — a model with tools, memory and a task loop, running on a schedule, doing work without a person in the conversation. Tell us about one you built and shipped. What it did, what it got wrong, how you found out, and what you changed. We are as interested in the failure handling as in the capability.
The discipline that makes a model safe to leave running
Some of what you build here touches federally regulated processes. Getting an answer mostly right is not good enough, and neither is escalating everything and calling it safe.
What separates an automation we can trust from a demo is the scaffolding around the model, and we have strong views on it:

• Deterministic checks bracket every model step. Validate the inputs before the model sees them, and validate the output against rules that do not involve a model. The reasoning sits in the middle, and it never certifies its own work — a verdict on correctness comes from a check the model cannot influence.
• Missing is a failure, not a pass. An empty result, a partial record, a field that did not populate — these fail loudly. The most expensive bugs in this class of system are the ones where nothing appeared to go wrong.
• Every output traces to the exact inputs that produced it. When a number is questioned six weeks later, you can show which records, which version of the logic, and which run produced it.
• Confidence thresholds are explicit. What the system handles alone, what it flags for a person, and what it refuses to decide — decided deliberately and written down, not emergent.
  You do not need to have used those words. You do need to have learned the lesson some other way, and to recognize why it matters more here than in most places. If your instinct on finding compliance logic in an unversioned script is that it needs fixing before it needs extending, you will do well here.
  Our stack — ideally you know some of this
• Python or Node for the pipelines. We are not precious about which
• Agent harnesses — the orchestration layer around the model, not just the model. Claude Code and similar coding agents, especially used across a team; custom harnesses you have built yourself; anything where you have had to decide what the loop does, where it stops, and what happens when a step fails. This is the part of the stack we care most about
• Agent frameworks and tool protocols — MCP, function calling, whatever you have used to give a model real capability
• Browser automation — some of what we need to reach has no API
• SQL and analytical databases — our data platform is ClickHouse
• ERP data — SAP Business One, though you do not need to have worked in it
• Source control and CI as habits, not as process imposed on you

Nice to have

• Pharmaceutical, medical or another regulated distribution environment
• Experience automating finance operations — AP matching, payment processing, reconciliation
• Document and data extraction from inconsistent third-party formats
• Having built something that a non-technical team then used every day without you

The engagement

• Six months, with a review at month four and the option to extend
• Full-time for the six months, as an independent contractor
• $50–75 per hour (USD), depending on depth of relevant experience
• Fully remote from Latin America, with at least four hours of daily overlap with US Eastern time. Our compliance and operations people are in New Jersey and the processes you are automating happen in their mornings
• Professional working English. You will be interviewing the people whose work you are changing, not just reading specifications