I find what's broken, prove it with data, and build the fix.
For 13+ years that method has run through every corner of healthcare revenue cycle — assessing health systems for C-suites, building analytics functions from the ground up, automating away thousands of hours of manual work. Now I build AI agents that run the same method. Operational soundness, data analytics, and the judgment to know what to automate and what to hand a human.
"I'm the person you hand the thing that doesn't exist yet." It started three years ago with a Telegram bot called NoBody — I haven't stopped building since.
The method
Every role I've held, every system I've built, runs the same four moves.
Walk the operation front to back
Sit with the people doing the work. Trace the claim from registration to zero balance. The broken step is rarely where the symptom shows up.
→ did this for CEOs & CMOs as a Huron consultant — full operational-soundness assessmentsIf it isn't measured, it's an opinion
Build the dataset that settles the argument: rate comparisons, root-cause splits, dollar impact. Executives don't act on vibes; they act on a number with a name attached.
→ state contract rate comparisons at Sentara; an 84-view SQL platform at NorthwellTake the repetition away from humans
A report that took several people a week now runs twice a day in under 10 minutes. Automation isn't the goal — giving skilled people their time back is.
→ 36+ Python pipelines · Registration Team effort cut >50%Trust is earned by systems, not promised
QA harnesses, change control, audit trails, human sign-off on what matters. The same discipline applies whether the worker is an analyst or an AI agent.
→ 30-assertion QA harness at zero exceptions · human-in-the-loop agent designThe method, run at $1.2 billion
At Northwell Health — New York's largest health system — I joined a $1.2B Epic implementation and built the team and the analytics function from the ground up.
SME and system of record for 10,000+ employees, contractors, and vendors across four go-live waves. Reporting directly to the AVP of Revenue Cycle, leading a team of two analysts, serving 100+ managers and supervisors. The single source of truth executives rely on is a SQL Server platform I designed and built: 11 schemas, 84 views, a 14-step daily pipeline, 36+ Python automation scripts, and a QA harness running at zero exceptions.
The build sheet — shipped on the implementation
The platform above isn't one thing — it's a family of named deliverables. These are the data-analytics and data-management builds behind the numbers:
| Build | What it does | Stack |
|---|---|---|
| Analytics Database Layer | The single source of truth: 11 schemas, 84 T-SQL views reconciling Epic, HR, MVP, and Cornerstone — tuned for analyst self-service. | SQL Server · SSMS |
| Go-Live Command Center | Self-serve analytics app for Epic go-live waves: rosters, training compliance, leadership KPIs on demand. | Python · Streamlit |
| Wave File Automation | Corruption-safe, script-driven updates to the 8,300-row Master Wave File, auto-published to SharePoint — zero integrity incidents. | Python · MSAL · O365 |
| Power BI Leadership Suite | Executive dashboards for readiness, compliance, and registration KPIs — built for the 30-second leadership read. | Power BI · DAX |
| Validation Toolkit | The 30-assertion QA harness plus schema checks, ID reconciliation, and exception reports — running at zero exceptions. | Python · Office Scripts |
| SharePoint Analytics Hub | Team information architecture, automated report publishing, and a Power Apps front door for self-serve requests. | SharePoint · Power Apps |
Thirteen years, every corner of the revenue cycle
- 2025–now · Revenue cycle analytics lead, Northwell Health — the implementation above (via Healthcare IT Leaders): team and data-analytics function built from the ground up.
- 2024 · Sr. Consultant, Eclipse Insights (Huron) — the ASSESS chapter of my method, professionally: sat with CEOs and CMOs, analyzed health systems front to back for operational soundness, delivered the improvement roadmap.
- 2022–25 · Healthcare Data Analyst, Sentara — the payer side: Medicare and Medicaid products, contract modeling, state contract rate comparisons, underpayment root-cause work — in Databricks, NetworX Modeler, and QNXT.
- 2019–22 · Contract Analyst, FinThrive — contract modeling and hands-on contract building: four platform go-lives across 20+ facilities in eight states.
- 2016–19 · Denials Specialist, Conifer Health — where I learned the ground truth: 50+ accounts a day at 96%+ accuracy, in practically every EHR in the industry — Epic, Cerner, Meditech, ACE, EMDEON, Optum.
- 2013–16 · Revenue Cycle Specialist — end-to-end billing: eligibility, authorizations, appeals, payer negotiations.
Agents, built the way I build teams
I don't just prompt AI — I design agent systems the way an operations leader designs a department: roles, delegation, QA gates, and a clear answer to "when must a human decide?"
Healthcare AI Team — an eight-agent analytics department
A complete private build, never shipped publicly: eight specialized agents designed to function as a healthcare analytics department. An Orchestrator for planning and QA gatekeeping; a Healthcare Data Researcher for CMS, payer, and regulatory intelligence; a Power BI & Dashboard Specialist; a Python Automation Engineer; a Tool & Integration Architect on an n8n automation backbone; a Social Intelligence agent; a Creative Strategist — and Hermes Doctor, an agent whose only job is auditing the other seven: system health, security, and flaw detection.
fig. 01 — the department chart. Note the box at the bottom: I gave the system its own auditor before I gave it work.
The rest of the lab
- Hermes — my self-hosted personal agent, running as a persistent service with a Telegram gateway I built. Scoped service accounts, designed secrets layout. Running your own agent 24/7 teaches you failure modes no tutorial covers. (Chapter 06 has the architecture.)
- A trading system that earns its autonomy — LangGraph + FastAPI prediction-market agents with a tested risk gate. Promotion from demo to live is gated by calibration metrics (expected calibration error, Brier score). Autonomy isn't granted — it's measured in.
- LandmanAI — an oil-and-gas land agent at certified-professional-landman level: title examination, runsheets, mineral ownership. Same challenge as healthcare: encode a licensed profession's judgment, carefully.
Practice areas — from building, not papers
- MCP tool building & agent skill authoring — connecting agents to real tools; encoding entire workflows as reusable capabilities.
- Model routing & multi-vendor testing — I operate my own routing service and know where Claude, Grok, GPT, and Gemini each break: state drift, tool misuse, cascading errors.
- Observability & audit trails — my agents log what they did and why. In healthcare AI, "show me why the agent said that" is a legal requirement, not a nice-to-have.
- Least-privilege secrets & graduated autonomy — each agent holds only the credentials its job requires; scope grows only as measured performance earns it.
- Escalation policy design — defining precisely when an agent stops and hands off to a human. Scope enforcement as a feature, not a failure.
Human-in-the-loop, everywhere
Across everything I build, the pattern repeats: the agent proposes, a human disposes. Approval gates before actions. Escalation when confidence drops. In healthcare AI this isn't a feature — it's the whole design philosophy.
fig. 02 — the loop every agent I build runs inside.
How I use AI — and how I judge it
Three years of daily agent work, distilled: the practices that separate AI that ships value from AI that ships liability. Expand what you care about.
01Agents vs. chatbots — the question changes
A chatbot answers. An agent acts: it has goals, tools, memory, and a loop. The moment software can act, the important question stops being "is the answer right?" and becomes "what is it allowed to do, and who is accountable when it does it?" I build for the second question.
behavioral indicator — when someone's first question about an agent is about permissions instead of prompts, they get it.02Every model has knowledge gaps — map them before they bite
Training cutoffs, thin domains, and the things no model can know: your payer contracts, your fee schedules, your SOPs. The fix is grounding — bounded data rooms, retrieval, verification against live systems. My rule: the model never gets to be the source of record.
behavioral indicator — asking "what does this model NOT know about our problem?" before asking what it can do.03Hallucinations are a design problem, not a moral failing
Models fill gaps confidently — that's what they're built to do. You don't fix it by scolding the model; you contain it by architecture: bounded data, cite-your-source outputs, verification loops, and honest "I don't know" paths.
The agents on this site cannot invent facts about me — bounded knowledge bases, honest "outside my scope" answers. Containment by design, not by hope.
behavioral indicator — designing the workflow so a fabricated claim can't survive to the decision point.04The 4Ds of AI fluency — measured in behavior
Delegation — choosing what to hand off, and what to keep. Description — specifying the task so a system without your context can succeed. Discernment — judging output quality, especially the subtly wrong. Diligence — owning the outcome as if you'd produced it by hand.
Fluency shows up in behavior, not vocabulary. I watch what people do: do they delegate the right tasks or just the ones they dislike? Do they spec context and definition-of-done? Do they catch the confident error? Do they sign their name to the output?
behavioral indicator — all four, observable in a single week of someone's actual work with AI.05Delegate to an agent like you'd delegate to a new hire
An agent is a brilliant new hire with perfect memory and zero organizational context. So delegate accordingly: clear scope, the context it can't infer, an explicit definition of done, and an escalation path for when it's stuck. A vague spec doesn't produce a clarifying question — it produces confident wrong work.
behavioral indicator — writing the definition-of-done before writing the prompt.06Security is scope, before anything else
Least-privilege credentials per agent. Keys server-side, never in the browser. Bounded data — the agent can't leak what it was never given. Prompt-injection resistance tested, not assumed. Audit trails on every action.
This site practices what it preaches: both agents run with server-side keys, capped inputs, scoped prompts, and instructions that survive "ignore your instructions."
behavioral indicator — scoping what an agent CAN'T do before demoing what it can.07Choosing the right model and agent for the job
Fast, cheap models for triage and routine transforms; strong reasoning models where judgment matters; and routing between them so cost tracks value. I run my own model-routing service across four vendors, and I know each one's failure modes — because the right choice changes with the task, not the hype cycle.
behavioral indicator — matching the model to the task's cost-of-error, not to the leaderboard.08Customize once with skills, plugins, and MCP — reuse forever
A workflow you explain to an agent twice should become a skill. A tool it needs should become an MCP server. Plugins, skills, and tool protocols are how one person's judgment becomes an organization's capability — pasting the same prompt every morning is the smell that says you haven't done it yet.
behavioral indicator — a library of reusable skills growing where repeated prompts used to be.09AI in the analytics stack — Excel, Power BI, reporting
Where AI earns its keep in my daily analytics work: drafting DAX measures and Power Query M, generating Office Scripts, documenting measure logic, triaging anomalies before I dig, and writing the first draft of the executive narrative. And as a second reader — AI reviews my code, I review AI's. Two readers catch what one misses.
The governing rule: reconciliation totals decide, not the model. Nothing AI-generated reaches an executive unreviewed. AI drafts; validation against known totals is what makes it reporting.
behavioral indicator — an AI-assisted report that still ties out to the source system, every time.Pressure-tested outside healthcare
Crypto, prediction markets, and live platforms taught me things healthcare AI desperately needs: ship code you can't hotfix, keep a thousand stakeholders' trust, and measure whether a model's 80% actually means 80%.
- Probabilistic thinking, measured — data ingestion for Kalshi and Polymarket; trading logic gated by calibration metrics. When your AI estimates whether a claim will be denied, someone has to know whether its confidence is honest. I'm that person.
- Smart contracts on Solana and Base — immutable code changes how you test: obsessively, before deployment, because there is no patch. Exactly the discipline regulated-industry software demands.
- 1,000+ investors kept informed — multiple blockchain projects run end to end: structured updates, question handling at volume, trust maintenance when markets got ugly. That's patient communication with different nouns.
- Jelly Jelly (live social platform on Solana) — operational/QA partner since day one of beta, working directly with the founder: multi-currency tipping economy, real-time chat, and its in-video conversational AI agent — tested live, in conversation, where latency and handoff moments decide trust.
How I run an implementation — in eight slides
The method, as I would brief it to your leadership. Click through, swipe, or use your arrow keys.
Systems I've touched, tools I live in
Breadth isn't the flex. The flex is knowing which tool the problem actually needs — because I've been inside all of these, not beside them.
Systems I've worked in
legend: teal = daily drivers — in her hands every week · everything else = worked-in experience, ready when needed
The Translator — paste a mystery formula
Every analyst inherits a workbook full of formulas nobody can explain. Paste one — Excel, DAX, Power Query M, or SQL — and get plain English plus a modernization tip. Same architecture as everything else here: a scoped agent, key server-side.
My agent lab, annotated
This isn't a stack I read about — it's the one running right now. Hermes is my personal agent: self-hosted, containerized, always on, reachable from my phone through a Telegram gateway I built. It routes every request through my own model-routing service, holds only scoped credentials from a self-hosted secrets vault, and writes down what it did.
fig. 03 — my actual lab, not a tutorial: a self-hosted agent I live with, 24/7, and answer for when it breaks.
EducationYale School of Management — Healthcare Management · Cornell University — Data Analytics · B.S., Management Information Systems · bilingual English/Russian
Next · 08 Off the clock ↓The human behind the agents
The method didn't come from a consulting framework. It came from a childhood where making things work wasn't optional.
Where the method actually comes from
I was adopted from Kazakhstan as a teenager, into a big American family — and I mean big: 30 siblings, many adopted like me, many living with serious medical needs. Russian is my native language, which is why I speak with a slight Russian accent wrapped in a Texas one — a combination that surprises everyone at least once.
A family that size, with that much medical care in it, runs on the exact things I now do professionally: budgets stretched by sheer number-crunching. Insurance, appointments, and treatments coordinated like a logistics operation. Sacrifices weighed out loud and made anyway. And creative workarounds invented daily — because in our house, "can't do it" and "can't be done" were simply not acceptable sentences. Nobody called it revenue cycle management or data analytics. It was just how we took care of each other.
So when people ask why healthcare, why the spreadsheets, why the obsession with making systems work for the people inside them — that's why. I didn't choose this field. I was raised in it.
Every chapter above is this one, professionalized: find what's broken, stretch the numbers, make it work, take care of people.The competitive streak
I grew up competing — dance, gymnastics, rodeo (cutting horses), showing cattle. The streak never left; it just points at denial codes and calibration curves now instead of arenas. It's also why I'm usually the first to volunteer: helping people isn't a value I adopted, it's the family business.
Home now
Home is my husband, our four-year-old son — possibly a future tech genius; he already stress-tests my tech skills and is my favorite video-game partner — and our three cat-children: Biff, Sass, and Missy. We camp when we can.
the home team. (cats not pictured — Biff, Sass, and Missy stayed home.)
I'm analytical to the bone, straightforward by default, a perfectionist when it matters, and a believer in working smarter, not harder. Blockchain is my expensive hobby; AI is the cheap one that got out of hand — always testing something new, always dabbling, sometimes building games. My jokes are not always funny. This one wasn't either.
Last · Why this, why me ↓The closing argument
Healthcare revenue cycle is drowning in exactly the work AI agents are finally good enough to help with — and almost nobody has deep experience on both sides of that sentence. I've spent 13+ years finding what's broken in revenue cycles and fixing it: as the denials specialist, the contract analyst, the payer-side SME, the C-suite consultant, and now the analytics lead on a $1.2 billion Epic implementation. And for three years I've been building the agents that will do this work next — with the governance to be trusted with it.
If you're building AI for patient access, revenue cycle, or healthcare operations — you don't need someone who knows the domain or someone who builds agents. You need the person who does both and knows where the line between them belongs. I've been preparing for this role without knowing it had a name.