Digion

Tyler Horan · Digion Inc.

I build and ship production AI systems.

Recent work: the backend and voice model for an AI sales coach deployed at Microsoft, a university-wide learning platform at Harvard, and clinical documentation for hospice care that comes out audit-ready under HIPAA. A decade of production engineering before a doctorate in computational social science — and nobody between you and the person writing the code.

● Available for contract and fractional engagements

Remote · New York metro

Selected work

Systems where getting it wrong had a name attached.

Some clients are named here with permission; others are described by sector.

Microsoft

Voice-driven AI sales coach

Scope
Backend and voice model for an AI coaching system that simulates expert sales technique, deployed to Microsoft sales teams.
Constraint
Coaching a subjective skill makes evaluation the hard problem rather than generation — a model that sounds right and grades wrong is worse than no coach at all.

Python · JavaScript · voice models

Harvard University

Learning experience platform

Scope
Principal engineer on a next-generation learning platform built for use across the university.
Constraint
University-wide means heterogeneous departments, existing systems of record, and accessibility obligations that aren't negotiable.

JavaScript

PhiliaCare

Hospice documentation that comes out audit-ready

Scope
Clinical platform turning hospice visits into notes, forms, and Medicare documentation automatically.
Constraint
Patient records in scope from the first commit. No third-party model exposure permitted; BAA and data-handling terms negotiated directly.
Outcome
Shipped with the data posture settled before it got expensive to change. Ongoing engagement.

Self-hosted open-weight models · private infrastructure

Environmental consulting

Vision-language extraction from historical maps

Scope
Structured land-use data pulled from scanned Sanborn fire insurance maps and USGS topographic quads, feeding Phase I environmental site assessments.
Constraint
Output carries legal weight. Run-to-run and cross-year consistency are the hard part, and human review stays in the loop by design rather than as a stopgap.
Status
In scoping with the delivery partner, estimated per document package.

Vision-language models · document pipelines

Employment technology

Independent bias auditing of automated employment decision tools

Scope
Adverse-impact audits of hiring and promotion systems under New York City Local Law 144.
Constraint
Statutory methodology, auditor independence, publicly posted results.
Outcome
Audit practice run under the Paritas name; available for independent engagements.

Disparate-impact analysis · audit methodology

Behavioral health & dermatology

Self-hosted inference inside the covered boundary

Scope
Intake and matching tooling for clinical placement, plus ongoing AI engineering on a clinician-facing dermatology product in practice settings.
Constraint
Sensitive intake material that could not reach a commercial endpoint; clinical accuracy expectations alongside patient data handling.
Outcome
Full stack on managed infrastructure with inference inside the BAA boundary. Continuing engagements.

Vertex AI · open-weight models under BAA

How I think about this

Most AI systems can do anything. The ones worth building can't.

Data that isn't allowed to move. Decisions someone can be made to justify. Outputs that end up in front of an auditor, a regulator, or opposing counsel. The engineering is the easy half.

Where constraints attach to an AI pipeline A four-stage pipeline — data, model, decision, people — with the regulatory and contractual constraints that bind at each stage listed beneath it. What you are building Data Model Decision People affected Where the rules attach Residency and retention Business associate agreements Consent and minimization Training and reuse terms Provenance and versioning Subcontractor exposure Bias and adverse impact Explainability of outcome Logging and reproducibility Notice and disclosure Recourse and appeal Liability when it's wrong
Every consequential system is defined by what it isn't allowed to do. Knowing where the constraint attaches — and building so it holds without crippling the thing — is most of the work.

Engagements

Three ways this usually starts.

Fixed scope where the work is knowable, retained where it isn't.

Build

AI engineering

End-to-end delivery on production systems: retrieval over messy internal corpora, agent workflows that actually complete, fine-tuning and evaluation harnesses, and the unglamorous integration work that decides whether any of it survives contact with real users. Where the data can't leave your environment — or the API bill has outgrown the convenience — this includes standing up open models on your own infrastructure, replacing third-party inference in systems already in production without a rewrite.

Project scope or monthly capacity · most engagements start here

Assess

Governance, audit, and exposure review

A fixed-fee review of where your AI systems meet the rules that bind them: vendor terms and what they actually permit, data exposure, model documentation, and adverse-impact testing where decisions affect people. Delivered as findings you can hand to counsel or a diligence process.

Fixed fee · scoped in the first call

Own

Fractional AI lead

Senior technical ownership for teams shipping AI without a senior machine learning voice in the room. Architecture, hands-on build, vendor decisions, and the risk calls that come with them — part-time, on defined scope rather than headcount.

Monthly retainer against agreed capacity

Who you'd be working with

One senior engineer, not an agency bench.

Digion is Tyler Horan. I spent a decade as a software engineer — lead and staff roles at Handshake, FairClaims, and Bright Software — before finishing a doctorate in computational social science, and I've spent the years since building AI systems for organizations where the output has consequences for somebody.

That combination is the point. Most people who can put a model into production have never read the agreement governing the data going into it, and most people fluent in that language have never shipped anything. The problems worth paying for live where those meet, and they don't get solved by one group writing findings for another.

I work directly with founders and engineering leads — no account managers, no handoff to a junior team. Agent-assisted workflows are part of how I deliver, disclosed to every client, and accountability for what ships is mine either way.

Doctorate
Computational social science, The New School for Social Research
Teaching
Lecturer, Data Analytics and Computational Social Science, UMass Amherst
Governance
ForHumanity — independent AI audit and assurance
Board
Director, Creative Capital Foundation
Writing
Three books under contract on technology, work, and attention

Start here

Tell me what you're building and what it isn't allowed to do.

Thirty minutes, mostly questions about your stack, your timeline, and your obligations. You'll leave with a scope and a number. If it isn't a fit, I'll say so and point you somewhere better.

tyler@digion.cloud