Agentic AI Engineer (Clinical AI)

InnovoCommerce; Irvine (CA); On-site; USD 180,000 - 240,000; Full time

Job summary

InnovoCommerce LLC in Irvine, CA is building agentic AI runtimes to work with clinical study teams. We are hiring an Agentic AI Engineer to own the agent loop, context management, and tool adapters across multiple providers including self-hosted endpoints.

The role requires strong Python, async, PostgreSQL, and familiarity with clinical data standards. You will design integrations, ensure safety, and help push the product toward reliable, compliant automation in clinical trials.

Qualifications

• Has built and operated a production agent loop from primitives, including tool dispatch, termination, retries, deadlines, idempotency, and mid-run recovery
• Context engineering at the mechanics level: token budgets relative to a model's window, prompt caching, compaction, and long-running conversation state
• Has shipped the same agent against more than one model provider and understands what a gateway does not carry across (reasoning blocks, tool-call formats, constrained decoding, finish-reason semantics)
• Evaluation-first habit: traces as the contract, deterministic scorers over judges where possible, regression gates before release, and comfort saying not measured yet instead of asserting
• Tool and skill design: registries, MCP, bounded sub-agents, and approval flows for non-idempotent actions
• Security posture for agent runtimes: prompt injection, least privilege for subprocesses, read-only data access for model-authored queries, redaction
• Strong Python with async, PostgreSQL, and gRPC or other typed contracts; able to read C#
• Clinical standards literacy or demonstrated ability to acquire it fast
• Deep commitment to the role and mission in building a product that fundamentally transforms clinical trials

Responsibilities

• Own the agent loop end-to-end: tool dispatch, limits, termination, retries, recovery, and event log.
• Design per-turn context: token budgets, prompt cache, in-loop compaction, and value-vs-reference decisions.
• Build provider layer: neutral request/response with adapters for Anthropic, OpenAI-compatible, and self-hosted endpoints.
• Define agents as configuration: manifests, capabilities, loadable skills, and a tool registry with approvals.
• Ground agents in clinical standards (CDISC) by wiring into the loop as active context and tools.
• Build evaluation nets: trace attributes, goldens, deterministic scorers, regression gates for tool choice/output.
• Harden the runtime: prompt injection paths, least-privilege sub-agents, redaction of regulated content from logs.
• Commute to the office for in-person collaboration (SoCal); discussions welcome for Bay Area.

Skills

Python

Async programming

PostgreSQL

C#

CDISC literacy

Clinical data standards

Tools

gRPC

REST/Endpoints

vLLM/LiteLLM

Job description

InnovoCommerce is building the next generation of clinical solutions: purpose-built agentic AI over a unified data foundation to work alongside study teams across the global life sciences industry. We’re seeking an Agentic AI Engineer with a startup mindset who’s ready to build the harness that runs our clinical AI agents: the agent loop, context management, tool layer, provider adapters, and the evaluation gates around all of it. This is a builder role for someone who’s shipped agent runtimes from primitives and solved the clinical data standards weeds OR can learn them fast and turn them into working agent behavior. We’re a motivated team that moves fast - come join!

Key Responsibilities:

• Own the agent loop end-to-end: tool dispatch, run limits, termination, retries, recovery after mid-run failure, and the append-only event log the loop runs on
• Design how context is built each turn: token budgets, prompt cache layout, in-loop compaction, and what enters by reference versus by value
• Build and maintain the provider layer: one neutral request and response shape with adapters for Anthropic, OpenAI-compatible, and self-hosted (vLLM/LiteLLM) endpoints, and close behavior gaps
• Define agents as configuration: manifests, capabilities, loadable skills, and a tool registry with effect classes, policy checks, and approvals for side-effecting actions
• Ground agents in clinical standards (e.g. CDISC) by wiring them into the loop as active context and tools
• Build the evaluation nets that gate every harness and model change: trace attributes, goldens, deterministic scorers, and regression checks on tool choice and output quality
• Harden the runtime: prompt injection paths, least-privilege sub-agents, and redaction of regulated content from logs and telemetry
• Commute to the office when necessary for in-person collaboration (SoCal location, but if you're in the Bay Area let's discuss)

Qualifications:

• Has built and operated a production agent loop from primitives, including tool dispatch, termination, retries, deadlines, idempotency, and mid-run recovery
• Context engineering at the mechanics level: token budgets relative to a model's window, prompt caching, compaction, and long-running conversation state
• Has shipped the same agent against more than one model provider and understands what a gateway does not carry across (reasoning blocks, tool-call formats, constrained decoding, finish-reason semantics)
• Evaluation-first habit: traces as the contract, deterministic scorers over judges where possible, regression gates before release, and comfort saying "not measured yet" instead of asserting
• Tool and skill design: registries, MCP, bounded sub-agents, and approval flows for non-idempotent actions
• Security posture for agent runtimes: prompt injection, least privilege for subprocesses, read-only data access for model-authored queries, redaction
• Strong Python with async, PostgreSQL, and gRPC or other typed contracts; able to read C#
• Clinical standards literacy or demonstrated ability to acquire it fast
• Deep commitment to the role and mission in building a product that fundamentally transforms clinical trials

Preferred Qualifications:

• Heavily Preferred: hands-on CDISC (clinical digital data standards/flow) or risk-based monitoring experience - if you have this, DM Sachil directly
• EDC, CTMS, eTMF, or other clinical system model exposure
• Regulated software background (e.g. GxP, 21 CFR Part 11)
• Experiential memory or self-improving agent systems distilled from traces, with awareness of reward hacking in closed evaluation loops
• Has authored evaluation cases directly from primary standards documents without waiting on subject matter experts