Trevor J. Romack

Technical project manager on healthcare data implementations, with eight years of professional experience. The last two have focused on client data pipelines: gap analysis and referential-integrity checks on inbound files, field mapping and validation with client and internal data teams, and the move from one-off extracts to recurring automated feeds. Alongside that I build data and AI systems — retrieval and grounding, evaluation harnesses, and audit logging among them — each documented end to end with its tests, evaluation results, and limits.

1. Claritev — Chicago, IL

   1. IT Project Manager, Associate

      Jun 2024 – Present

      • Inbound client data, mapping & acceptance: Work incoming health-plan data files alongside client and internal data teams — comparing what was requested against what arrived, identifying null and malformed fields, and confirming records reconcile back to the member or claim they belong to. Coordinate mapping between client source systems and our data model, then drive the validation and client acceptance cycle a program must pass before go-live.
      • Automated feeds & data-quality exceptions: Support the move from one-off extracts to recurring automated feeds, then work the exceptions that follow — schema drift, unannounced code-set changes, and volume anomalies that usually indicate an upstream system problem. Escalate and resolve with the client’s IT group when their extract stops matching their own system of record.
      • Regulated data handling: Operate under HIPAA-aligned requirements for member and claims data throughout, coordinating with compliance and security before data moves.
      • Delivery ownership: Lead cross-functional delivery of post-payment implementations (Subrogation, Coordination of Benefits, Data Mining, MSP, ESRD) across concurrent client programs — scope, schedule, dependencies, risk, and status reporting to client and executive audiences.
      • Internal tooling built: Built a Project Management Dashboard end to end with Claude Code, unifying delivery status previously scattered across Jira/Confluence, Salesforce, SharePoint and email; adopted by peers beyond my own team. Also designed an AI-assisted Milestone Generator that standardizes implementation timelines — Claritev AI Competition Winner, December 2025.
      • AI adoption: Defined what qualifies a user at each of three Claude tiers, audited Anthropic’s course catalogue against those competencies, and designed the resulting curriculum — approved as proposed and still the recommended path company-wide, for a population of ~500 AI-tool licenses.

   2. Case Coordinator

      May 2022 – Jun 2024

      • Analyzed healthcare claims and recovery data across subrogation workflows — reconciling records, investigating exceptions, and supporting audits, recovery planning, and compliance reporting.
      • Partnered with internal operations and client contacts to resolve discrepancies between internal records and client-supplied data, and to document findings.

2. Luna Flooring — Chicago, IL

   1. Project Manager / Sales Representative

      Mar 2019 – May 2022

      • Owned residential projects end to end — consultation, scoping, planning, execution — delivering on scope, timeline, and budget while managing client relationships and coordinating crews and vendors.

3. Discovery Health Partners — Itasca, IL

   1. Subrogation Specialist / Investigations Specialist

      Apr 2018 – Mar 2019

      • Validated and managed healthcare data for legal and financial recovery matters; investigated insurance claims and member data, collaborating with attorneys and insurers and communicating findings to stakeholders.

Built and shipped

1. Project Suver — Governed multi-tool AI product

   • Case study →
   • Live demo (may take ~30s to wake) ↗
   • GitHub ↗
   • One shell, sixteen tools: 16 tools across 4 platforms on one reusable no-prompt shell over five vendored engine cores — a sanitization boundary, a summarizer, a template drafter, a typed extractor, a two-source comparator. A new vertical is a configuration — a field set, a draft kind, a policy — not a fork.
   • Grounding & privacy: every answer cites the source span it came from or the tool abstains, and a draft section that cannot be grounded is withheld rather than written. PII/PHI is tokenized locally before any model call and re-hydrated only in the user’s view, so the boundary is not a step the model can skip.
   • Retrieval: ingestion and splitting, then lexical retrieval over stemmed token overlap widened by model-assisted query expansion. An optional LLM re-ranker reorders hits but never drops one, and ships off by default: enabling it fixed one retrieval miss and made a neighboring case answer a question it should have refused.
   • Evaluation: a labeled 20-case trust set scored on the real model across answerable, unanswerable, adversarial, and sensitive categories — 20/20, zero hallucinations, zero fabrications — plus a deliberately messy hard set at 7/9 with every miss root-caused and published. 225 automated tests run offline on a deterministic stub, so the suite needs no API key.
   • Stack: Python, FastAPI, Jinja, htmx, SQLite, Docker, pytest. Synthetic and public data only.

2. spancheck — Citation-grade evaluation for grounded-answer systems

   • Case study →
   • GitHub ↗
   • What it does: scores any RAG or retrieval-then-answer pipeline on citation accuracy, abstention correctness, hallucination rate, and cost and latency, then emits a versioned audit log built for a compliance reviewer. Installs from PyPI, runs as a three-command CLI, and drops into CI as a composite GitHub Action that fails a pull request on a regression.
   • Citation verification: each citation must clear two deterministic checks — provenance, that the cited span is verbatim in the retrieved context, and support, that the span covers the claim. A fabricated quote and a right-answer-wrong-citation both fail. Provenance is never delegated to a model.
   • Calibrated judge: the LLM-judge is scoped to entailment, the one check a rule cannot do, with prompts in version-controlled files and every verdict stamped with its prompt version and model. Agreement against a human gold set: 0.75 for the deterministic stub, 1.00 on claude-sonnet-5, no false positives either way.
   • Measured against a live system, twice: dogfooded black-box against a shipping RAG pipeline. On one tidy contract — 1.00 citation accuracy, 0 hallucinations, 0.967 overall across 12 cases — it caught a real false abstention, where the system declined an answerable question. On a deliberately self-conflicting four-document corpus, 21 cases scored 1.00 overall with 47 of 47 citations verified against the specific document each named, both sides of every conflict attributed correctly.
   • Proved on a system I did not build: pointed at Kotaemon, a published third-party citation-RAG application, over a public healthcare corpus — 18 of 18 verbatim citations source-scoped to the exact document named, with no tuning. It caught a fabricated citation, and exposed an assumption in my own support metric that I then fixed and shipped.
   • Stack: Python standard library only, zero hard dependencies, 72 tests green on Python 3.11 and 3.12. MIT-licensed, unlike the products it grew out of. Extracted from an earlier in-repo harness and generalised; the inheritance is recorded in the repository.

3. clear-pricer — Public hospital price and provider data, cleaned and queryable

   • Case study →
   • Data release ↗
   • GitHub ↗
   • Pipeline: ingests three hospitals’ CMS price files, the full NPPES provider registry and a synthetic FHIR population, validates and reconciles them behind quality gates, and republishes a clean dataset. 7,371,416 charge rows and 9,839,369 provider NPIs under type-2 change-data-capture, so re-applying a file changes nothing and a weekly delta touches only its own records. The same CLI steps run under Airflow locally and on a daily GitHub Actions schedule.
   • Gates that stop the run: dbt build runs 64 tests, every one able to fail the pipeline — a deliberately broken input turned the run red, publish never ran, and the served database kept its previous build. Every publish also checks counts, keys and money sums between the DuckDB warehouse and the served Postgres.
   • Findings published, not hidden: a 23-entry schema-drift log written as each deviation was found. 92.8% of one hospital’s payer rows report a median over zero claims, 55.9% of its dollar-plus-percentage rates do not reconcile against its own list price, and another publishes 39,116 rates at the full chargemaster price. Every price carries a field saying where its dollar came from.
   • Reproducibility and liveness: a GitHub runner and a Windows workstation independently built the same release fingerprint — 15 Parquet files, byte-identical — and pinning the generator’s end date made that hold across days, not just machines. Every release ships check values — row counts, distinct keys, the charges-to-codes fan-out, integer-cent checksums — that recompute from a clean download, alongside the 9.88M-row provider change history. A watchdog confirms the pipeline has succeeded within 26 hours; over the first week GitHub fired 21 of 21 scheduled events, each between two and a half and eight hours late rather than dropped.
   • Stack: Python, DuckDB, dbt, Airflow 3, Postgres, FastAPI, Docker, Parquet. 124 tests. Public data and synthetic FHIR only; no PHI.

4. biomed-evidence-mcp — Multi-source retrieval over MCP, with relational staging under it

   • Case study →
   • GitHub ↗
   • Architecture: six MCP servers expose ClinicalTrials.gov, four openFDA datasets, and PubMed to Claude as callable tools over one canonical Postgres table — 212,919 documents — so a cross-source question is a single SQL join rather than three API calls stitched together by a model.
   • Retrieval: hybrid search over pgvector cosine and Postgres full-text, fused with Reciprocal Rank Fusion; scored on 20 labelled questions at hit@1 70%, hit@3 80%, hit@10 95%, MRR 0.77, with every miss and its root cause published alongside the scores.
   • Relational staging beside the document store: documents answers what to cite; it has no grain and nothing joins, so it cannot answer how many distinct adverse-event cases there were as opposed to reports. A staging layer loads the same upstreams’ relational files into real tables — 29,497,843 rows across 14 tables, four FAERS quarters and a pinned AACT snapshot, partitioned so re-loading a quarter replaces that quarter alone.
   • Gates that prove the grain: five gates — grain, conservation, dedup, integrity, rejects — pass over 31,031,531 rows, each written so a clean database returns nothing. The grain gate disproved a key I had declared: 7,294,235 FAERS drug rows carry only 6,569,196 distinct (primaryid, drug_seq) pairs, so it is a join key and not an identity. The integrity gate’s 857 orphans were all explained by 62 rows the reject lane had already set aside.
   • Ingestion: per-source watermarks that advance only on success, backoff and caching against rate-limited public APIs, and browser TLS impersonation for a source behind bot management.
   • Verification: 433 tests, 182 of them contract or failure-path — the contract tests drive list_tools() on every server and assert each tool’s registered schema, so a client-visible change cannot pass unnoticed; clean-clone CI that boots the stack and runs a keyless search on every push; a 14-step end-to-end tour exercising all nine tools over real MCP stdio.
   • Stack: Python, PostgreSQL, pgvector, MCP, Docker, Voyage embeddings, Claude Haiku 4.5 for the three summarization tools. Public and synthetic data only; no clinical evaluation.

5. NBA Daily ETL — Daily Airflow and PySpark pipeline

   • Case study →
   • Live dashboard ↗
   • GitHub ↗
   • Pipeline: five-task Airflow DAG pulling NBA box scores nightly, transforming them in PySpark, and landing Parquet partitioned by (season, game_date) across raw, staging, processed, and feature zones on S3 or local disk. Ran unattended through the 2025–26 playoffs and Finals — 2,630 team-game rows across 208 game dates and 30 teams.
   • Correctness under reruns: staging-then-promote, so a transform that dies halfway leaves a partial staging directory rather than a half-written canonical history; dynamic partition overwrite, so a rerun replaces only the partitions it produced instead of the season.
   • Evaluation: strict date-boundary walk-forward cross-validation — no random k-fold on time-series data — reported against three named baselines on the same out-of-fold rows. Season-final logistic regression scored 0.6407 against the best baseline’s 0.6366 on 993 out-of-fold rows, and is described as inside the noise band rather than as an edge.
   • Verification: 160 tests passing in under two minutes with no AWS credentials and no network access; GitHub Actions runs ruff, black, pytest, and a Docker image build on every push. A 14-day playoff backfill completed 70 of 70 task instances with zero failures.
   • Stack: PySpark, Airflow, Parquet, AWS S3, Python, Streamlit, MLflow, Docker, pytest. Public data only.

6. Applied-AI Engine Suite

   • GitHub ↗
   • LLM Evaluation & Guardrails Harness — grades a system against golden and adversarial sets, gates regressions in CI, filters PII and hallucinations.
   • Payment-Integrity Claims Reviewer — rules/ML layer flags likely claim issues, a model explains each in plain English, a human approves or dismisses. Never auto-decides.
   • Agentic Workflow Orchestrator — plans a goal into steps with guardrails, a human checkpoint, and a full audit log.
   • AI Use-Case Intake & Prioritization Console — structured intake with scoring across five dimensions; flags where AI is a poor fit.

Hands-on — built and shipped with

Python · SQL · PostgreSQL · FastAPI · REST APIs · Docker · Git · SQLite · ChromaDB · pgvector · Pandas · PySpark · Airflow · Parquet · pytest · CI checks · RAG and retrieval pipelines · embeddings and chunking · prompt engineering · model evaluation / eval harnesses · agent orchestration · MCP servers · Claude and Claude Code

Applied in delivery — used in a coordination capacity

Data mapping and validation · reconciliation and referential-integrity checks · ETL/ELT concepts · schema and data-quality exception handling · HIPAA-aligned data handling · AWS (EC2, S3, Lambda, IAM) · Power BI · Tableau · Jira · Confluence · Salesforce

Project & delivery management

Full-lifecycle delivery · requirements elicitation and documentation · scope, schedule, and risk management · client onboarding and acceptance · status and executive reporting · stakeholder and vendor coordination · change management and enablement · Agile and SDLC

Domain

Healthcare payer operations · post-payment (subrogation, coordination of benefits, data mining, MSP, ESRD) · claims · HIPAA and compliance

• Loyola University Chicago — B.A., Digital Media Communications

  May 2016