Full-Stack Engineer — Healthcare Operations Automation

We build automations for healthcare operations. Things people used to do in spreadsheets, portals, and email — intake, reconciliation, reporting, review, workflow hand-offs — we replace with small, focused applications that staff actually use every day. Today it's a one-person function. You would be the second.

The work has three things in common across every project:

• LLM- and agent-powered. Most new builds use LLMs for extraction, classification, rule evaluation, or multi-step reasoning. You should be comfortable designing agentic workflows — tool use, structured outputs, evaluation loops, cost and latency awareness — not just calling an API and pasting the response.
• Deployed somewhere real. Cloud (Azure is the default) or on‑prem / local, depending on the customer. You own the packaging, the install, the update path, and the feedback loop when something breaks at 7 AM.

• You’ll be on the front edge of how AI is actually used in healthcare operations. LLMs, agentic workflows, and coding agents are the default tools for new builds — not a roadmap item, not a pilot, not a slide. If you want to work where applied AI is the baseline, this is that environment.
• You’ll be surrounded by early adopters. The operators across the portfolio are willing to try things, give blunt feedback, and put new tools into their daily workflow within days of seeing them. That kind of feedback loop is rare and it makes you a better engineer fast.
• Internal tools can become real products. When something we build for the portfolio turns out to be good enough to serve a broader market, we have a clear path to spinning it out as a standalone, supported company. The work you do here doesn’t have to end its life as an internal tool. That option exists, and it’s part of why this seat is worth taking seriously.

Representative examples, not a Wishlist:

• Take a workflow that currently lives in Excel and manual portal lookups and turn it into a web app a non‑technical user can run themselves. Build the ingestion, the rules, the UI, and ship it.
• Design an LLM‑based extraction pipeline for a messy document type (PDFs, scans, multi‑format exports) with structured outputs, confidence scoring, and an exception queue. Make it reliable enough for production.
• Stand up an agentic workflow — a chain of LLM calls and tool invocations — that automates something a human analyst used to do in 40 clicks. Write the evaluations so we know when it regresses.
• Own deployment for a specific app: container build, Azure Container Apps / App Service setup, SSO, logs, alerts. Or the equivalent on‑prem: signed executable, installer, update story.
• Pick up in‑flight projects and finish them. Some are 80% done and waiting for someone to close the last 20%.
• Write the spec, the README, and the handoff doc. Respond to stakeholder review notes in writing.

Most of these. Be honest about which you haven’t done.

• Shipped a full‑stack application used by real non‑technical users. Back‑end + front‑end + a deployment that someone other than you maintains.
• Strong Python. Comfortable with pandas and file I/O against messy real‑world data. Comfortable with at least one web framework — FastAPI, Flask, Streamlit, Django.
• Reasonable front‑end. Doesn’t have to be a React specialist, but should be able to build a usable interface — React / Type Script, or a Python‑first framework like Streamlit, or well‑structured plain HTML/JS.
• Called an LLM API in production code — Azure OpenAI, OpenAI, Anthropic, or equivalent. Structured JSON outputs, retries, cost awareness, basic evals.
• Understands agentic patterns in practice, not in theory. Tool use, multi‑step reasoning, when to use an agent vs a plain prompt, when NOT to use an LLM at all.
• Deployed something to a cloud provider (Azure preferred, AWS / GCP acceptable) with auth, logging, and a sensible CI/CD path.
• Packaged something for on‑prem or desktop use — Docker, PyInstaller, MSI, signed executables. Or has the judgment to learn it quickly.
• Can work on Windows. A meaningful share of our users and build targets are Windows‑first.