Jack Wright — AI / Automation Engineer

01 Shipped systems

Internal production systems. Described at the capability level; no client or employer internals.

Contract tooling for legal review

Not one system — a set of tools built around how the reviewer actually works, each with a different delivery vehicle.
Obligation reconciliation. Long-form agreement in; 500+ discrete obligations out, checked against the standard template into a coverage matrix with each gap routed to whoever owns it. 92% gap recall at 0% false positives on a hand-graded set with negative controls. ~150 tested rule functions keep the reasoning auditable; the model only does semantic matching.
Subcontract template builder. Generates the subcontract off the governing prime contract, adjusted into house language and standing preferences.
Prime contract first-pass review. In progress.
A shared workspace, not an app. For one engagement the right answer was a configured Claude project the lawyer works inside directly, alongside the model. Shipping software would have been slower and worse.

Python
Document intelligence
Calibration
Legal workflow

Analytics warehouse

Ingestion → dbt staging/marts → executive-facing views. First data substrate at the company.
Enforced dbt model contracts: a breaking schema change fails CI, not a dashboard.
Snapshot + milestone fact tables add history to sources that only store current state. Enables schedule-slippage analysis.
Prefect orchestration on a managed worker. Azure infrastructure from Bicep, including decommissioning an environment that wasn't earning its cost.

dbt
SQL
Prefect
Azure
Bicep

Lead-generation engine

Input: public parcel + assessment records. Output: ranked, contact-ready target list with per-row rationale.
Pipeline: building-spine filter → address-level firmographic matching → multi-gate qualification → buying-committee enrichment.
Delivered ~160 verified targets against a 150–200 ask. Adversarial verification removed ~40% of candidates as wrong or unverifiable before anyone made a call.
Config-driven: a new market is a profile file plus reference data, zero code change. 71 tests. Output ships as a living tracker that upserts on re-run and never overwrites human-entered columns.
In use — the leads are in the sales system, not in a spreadsheet somebody was emailed once.

Public records
Entity matching
Adversarial verification

Lesson-planning AI — Brilla Public Charter Schools

Mode 1: unit overview + lesson number → complete, teach-ready lesson plan grounded in the network's curriculum and teaching expectations.
Mode 2: multi-turn coaching that walks a teacher through internalizing a plan before class.
Core requirement: quote accuracy at scale. Every passage a plan cites is extracted verbatim from the purchased book text, not recalled by the model.
Extraction is a five-strategy parser cascade with guards against the silent failure — a parse that succeeds and is wrong. Chapters carry source offsets, so every quotation is traceable to a position in the book.
In deployment. 900 lessons — 9 grade levels × 4 units × 25 lessons — are the current school year's K–8 reading curriculum, not a pilot corpus. All 900 human-reviewed and passed before entering classrooms.
Book-specific plans are normally licensed from a publisher or commissioned from a curriculum consultant, which is why only well-resourced schools have them. This produced a year of them for a few hundred dollars of compute, and reruns for next year's books at the same cost.
Conditioned on grade level, not just grade band, so lessons are at-level: two grades inside one band reading the same text get different vocabulary load, scaffolding, and practice expectations.
Teacher adversarial review is the correction signal: reviewers attack each plan, and captured corrections improve the next generation run.
Per-mode model selection + prompt caching: ~$0.05 per generated plan; full-school usage projected at tens of dollars per month.

LLM generation
Text extraction
Quote grounding
Human feedback loop

Company-wide AI adoption

Administer the enterprise AI tooling the whole company uses, including access, governance, and spend monitoring.
Internal knowledge base wired into those tools, so retrieval happens where people already work rather than in a separate app.
Package domain expertise as installable skills and distribute them to non-technical teams — finance, preconstruction, marketing.
Train staff directly. Ran the platform evaluation and wrote the decision brief leadership used to pick a direction, and the acceptable-use groundwork alongside legal.
Adoption is the measure, not deployment: the contract engine is in active use by the reviewer it was built for.

Enablement
Governance
Cost monitoring
Training

Knowledge platform + MCP server

Retrieval-backed knowledge base: per-user row-level isolation plus a shared organizational tier.
Delivered inside staff members' existing AI assistants — no new app to open. Multiple pilot users live.
Fabrications on a fixed eval: 4 → 0, via an explicit grounding rule and identifier guards. Quality gate fails on omissions, not just inventions.
Shared MCP server exposes internal data and tooling as typed tools; stdio and hosted OAuth entrypoints.

RAG
MCP
Row-level security

02 Research & projects

Autonomous MTG Arena agent

Harness for LLM-driven play of Magic: The Gathering Arena. The game has no API, so the bot drives the real client through its interface; the client publishes the legal move set, so the model only ever chooses among options it is already allowed to take.
Live result: a naive Haiku pilot with no game-specific tuning played 55 games against Arena's built-in AI and finished 27–28, a 49.1% win rate, at ~$0.03 per game — while the actuation layer still had known defects.
Loop: perceive (game log as truth, vision only for screen position) → decide (typed request to the model) → execute (verified clicks). Every reported success is provisional until the engine confirms it.
Runs on a fleet of VMware VMs with a health sensor that classifies nodes CRASH / WEDGE / ANOMALY / CRAWL / HEALTHY.
Engineering practice: log-first diagnosis over vision, per-mechanic eval gates, adversarial multi-agent code review.

LLM agents
Computer vision
Python
Private repo

Architecture → The LLM pilot → Data research →

MTG analytics research

Mana-efficiency study over 4.7M games of 17Lands data: correlation between mana spent and win rate.
16 vs 17 lands in Limited; Curve Score deck metric; player-skill vs rank-distribution study.
Mana Bug: downloadable app that computes the metrics from your own Arena log files.

17Lands data
Statistical modeling

Research hub →

Narrative Tree Complexity Metric

With a psychology research group at the CUNY Graduate Center; co-author on the paper.
Method: expert pairwise comparisons of interactive-narrative trees → complexity hierarchy → supervised + unsupervised models that score new trees.

Research
ML modeling

Paper → Presentation →

Earlier work

Linear regression tool — Shiny app: upload a dataset, step through multiple/logistic regression.
NFL quarterback performance — Markov-chain simulation of career expected points added.
Google Gstore revenue prediction — Kaggle team project; technical writing sample.
Audio downsampling — sample-rate standardization to fix phasing in a Hammond B-3 virtual instrument.

03 Publications

Wright, J., Daiute, C., & Murray, J. T. (in preparation). Developing and refining the IDN-LLM-Autocoder. First author. Automates a hand-built narrative coding scheme — human annotation in ATLAS.ti was the throughput ceiling on study scale. Two independent classifiers (player reflections, IDN design elements) scored per category against a human gold standard; reliability targets Cohen's κ ≥ 0.70 / F1 ≥ 0.75, with low-recall categories fed back into taxonomy revision.
Daiute, C., Wright, J., & Murray, J. T. (2025). Streams of consciousness in interactive digital narrative research design. In Interactive Storytelling: ICIDS 2024. Springer. Springer →
Daiute, C., Murray, J. T., Wright, J., & Calistro, T. (2022). Intersubjective pivots in interactive digital narrative design learning. In Interactive Storytelling: ICIDS 2022, LNCS 13762, pp. 366–382. Springer. Springer →
Wright, J. J. (2022). IDN tree complexity (MS thesis). Graduate Center, City University of New York. Related report →

04 Writing

Mostly one question: how do you know an AI system is actually working?

Teaching a language model to play Magic

A personal project, and the one where the ideas get tested at full depth. Four pieces, in reading order.

Building a bot that plays Magic: The Gathering Arena
The game has no API, so the bot uses it the way a person does. How the game's own log made the vision layer almost disappear, and why a click that reports success usually didn't do anything.
An LLM refused to block 97% of the time
A naive pilot reached 49% against Arena's own AI for three cents a game, then declined to block on 268 of 276 chances. It wasn't bad at combat — power and toughness appeared in zero of 4,201 prompts.
How do you put a number on a Magic position?
A card's value can't be read off the card, every measurement of it is contaminated by the deck it came from, and what you condition on determines what the agent would learn to want. Some of the game can't be scored from a position at all.
Errors hiding in a 5.1M-game dataset
A card that isn't a card, shipped with an 18-point win-rate outlier. 23 cards that were never once cast but kept turning up on the battlefield. And a join proved correct by arithmetic instead of inspection.

Systems built at work

Described at the capability level — no employer internals, no client names.

Generating a school year's reading curriculum
Book-specific lesson plans are a premium product most schools can't afford. This made a year of them — 900 across nine grade levels, every quotation verbatim from the real book, every lesson attacked by a teacher first — and it reruns next year for the same cost.
Reading 500 obligations out of a contract
92% gap recall at zero false positives — and the rubric that reported a clean bill of health because it couldn't see the number that had doubled.
Finding sales targets in public property records
Parcel data to 160 contact-ready targets. Asking the model to refute rather than confirm killed 40% of candidates before anyone made a call.

05 Methods

Deterministic rules / model judgment split

Rules that must be auditable become tested code. Only genuinely semantic work goes to a model. Cuts cost and error rate.

Evaluation as a build gate

Graded answer keys, regression runs, negative controls. A change that fixes one case and breaks another is visible before it ships.

Adversarial verification

Independent refute-by-default reviewers over every candidate finding. Perspective-diverse verifiers beat redundant ones.

Deliver into existing tools

Ship into what the user already opens. End-to-end scope: substrate, model layer, interface, adoption.

06 Skills

AI / agent engineering

Applied LLMs
Agent architecture
Multi-agent verification
RAG & retrieval
Evaluation & calibration
MCP
Structured output

Data platform

Python
SQL
dbt
Data modeling
Prefect
Azure
Infrastructure as code

Document intelligence

Obligation extraction
Reconciliation logic
PDF / OCR forensics
Deterministic doc generation

Foundations

Statistical inference
Supervised & unsupervised learning
R
Requirements elicitation
Technical writing