RYAN ZERNACH

Full-Stack AI Systems Engineer

Ryan_Zernach_2025_Senior_AI_Systems_Engineer_Remote_United_States

BUILDING RELIABLE

HEALTHCARE AI

SCALING PRODUCTION

LLM SYSTEMS

ENGINEERING TRUSTED

AI INFRASTRUCTURE

DELIVERING INTELLIGENT

PATIENT IMPACT

⚡️ ZurboTax: AI Engineering for Federal Tax Preparation

Tax season is a poor time to reconstruct a year of financial decisions. Instead of using AI as a gimmick, I built a private tax-engineering system that ingests fragmented records, normalizes them into deterministic artifacts, computes tax workpapers, autofills official forms, and assembles a review-ready submission packet with field-level audit payloads. After 12 years of operating LLCs, it made this my fastest, cleanest, and most professional preparation cycle: the process finally behaved like software engineering, not seasonal spreadsheet triage.

Related Links
Finequities: SOC 2 Compliant Mobile App
Agentic federal tax preparation workflow image showing Codex, GitHub, AI, LLMs, and automations

Why This Is Memorable

Most AI tax stories stop at the prompt. This one starts with production discipline: validation gates, year-scoped filing directories, deterministic JSON and markdown artifacts, official-form autofill, and a packet build that fails loudly when something is missing.

Why This Is Important

The useful startup lesson is not “automate more.” It is ship quickly with hard guardrails, transparent automation, and outputs that still hold up when an operator, customer, or investor asks how they were produced.

What Makes It Different

Every critical output can be rerun, inspected, and explained. For AI systems that touch public-trust workflows, that is not a premium feature; it is the floor.

Technical Deep Dive: tax-filings

The production spine lives in my `tax-filings` repo, not in marketing copy. It is a year-aware filing system with explicit validation, deterministic outputs, and inspectable form payloads—the practical evidence needed to review, rerun, and explain the work.

Repository architecture and execution flow

Input normalization and correctness controls

Official form autofill and auditability

Submission packet generation and safety boundaries

Reliability mindset for high-consequence systems

Failure modes I engineered against

Summary

Taxes often arrive as an annual emergency. I treated them as a systems problem: replace panic cleanup with a repeatable engineering workflow that holds its quality under deadline pressure.

Execution Surface

A transaction-ingestion and classification engine feeds a year-scoped filing pipeline: validated inputs, normalized outputs, computed workpapers, official-form autofill, and a merged filing packet.

Outcome

The biggest win was operational credibility. Preparation became faster and cleaner because the workflow was inspectable, repeatable, and reusable instead of ad hoc.

Key Details

The sections below trace the original tax-prep pain, the finance system that addressed it, and why GitHub plus Codex made the workflow practical to sustain.

Why I built this in the first place

What the dashboard actually automated

How GitHub and Codex fit into the loop

What changed by tax season

What Moved Toward Autopilot

  • Multi-source transaction ingestion and normalization into one review surface.
  • Business-versus-personal classification persistence through manual overrides and reusable rules.
  • Tax-year-specific workpaper generation with reproducible JSON, markdown, and PDF artifacts.
  • Official form autofill with adjacent field payload audit logs.
  • Submission checklist and packet assembly that enforces required-form presence.

LLM Infrastructure Alignment

This is the engineering posture AI product teams need: deterministic pipelines, strict validation, transparent assumptions, and explicit human accountability for final decisions.

Principles I Enforced

  • Deterministic outputs over vibe-driven spreadsheets.
  • Human accountability over unchecked autonomy.
  • Auditability over black-box convenience.
  • Repeatable runbooks over heroics.
  • Measured operational discipline over one-time demo theater.

Full-Stack AI Systems Engineer

This is the operating style I bring as a Full-Stack AI Systems Engineer: move fast without getting reckless, automate without creating opaque risk, and ship systems that stay legible when the stakes rise.

  • I can design AI systems that operate under scrutiny, not just in demo environments.
  • I can translate ambiguous operational workflows into deterministic pipelines with explicit controls.
  • I can ship automation that accelerates throughput without surrendering accountability.
  • I can build systems whose outputs are inspectable, reproducible, and defensible in front of reviewers.