82W DigitalStart a conversation

82W Digital

Shipping AI into
places it usually
doesn't go.

The models work. Most deployments don't. 82W runs a phase-gated delivery discipline, with evals before the build and transfer at the end, for regulated-industry AI and for turning committed AI spend into shipped, measured value. Senior hands on every engagement, end to end.

Start a conversationHow we work →

What we do

The models work. The deployments don't. We close that gap in the places where the system has to hold up the first time.

The chasm between a promising pilot and a production system isn't a model problem. It's integration, quality at volume, observability, and ownership. A better model fixes none of those. A delivery method fixes all of them: define what working means before the build, gate every phase on a written exit condition, and hand over a system your team can run.

One senior practitioner from first call to handoff. No junior swap-outs, no staff augmentation. The work is architected to be handed off, not to create dependency.

Practice areas

Two practices. One way of working.

Everything we take on sits in one of two lanes. Both lean AI-first, both ship working code, and both run the same phase-gated discipline end to end, with the same senior hands.

Eval-defined. Phase-gated. Shipped.

AI Prototyping Studio

We build AI systems for places where it has to work the first time: clinical settings, behavioral systems, regulated workflows. The eval suite is the spec. Passing behavior gets defined, version-controlled, and agreed before the build, so non-determinism becomes a managed property instead of a standing risk.

  • Regulated-industry AI

    HIPAA-compliant builds, auditable pipelines, PHI-aware architectures. We've shipped into behavioral health and education environments where the rules are real.

  • Evals-as-engineering

    Golden trajectories built from your hardest real cases, regression across model versions, LLM-as-judge calibrated against human review. No eval suite, no build.

  • AI agent development

    Task-shaped agents, tool-use systems, and multi-step workflows, with constitutional guardrails, escalation paths, and review loops that keep humans in the decision. Built to be inspected, not just invoked.

  • Phase-gated delivery

    Validation through Transfer, with a written exit condition between every phase. Working code in real hands in weeks, and a system your team runs without us at the end.

Committed spend, shipped value.

AI Value Realization

Most enterprise AI budgets are already committed to model contracts, platform agreements, and copilot seats. Most of that spend hasn't converted into shipped, measured value. We close the gap between what you bought and what your leadership was promised.

  • Committed-spend audit

    A week-one inventory of the spend and the stalled initiatives against your actual data and systems. The output is a decision: the one use case with the shortest credible path to measurable production value.

  • Value-realization sprint

    One use case through all five gates: eval-defined, built, hardened, shipped into the real workflow. Proof, not another pilot.

  • Production AI on your platform

    Agents, RAG systems, and decision surfaces built on the data platform you already run. Tooled into your data, governed by your controls, deployed where your team already works.

  • The value memo

    Instrumented workflows, lift measured against the eval baseline, cost per outcome. Two pages your leadership can take to the board, and the artifact that defends next year's line item.

Selected work

Things we've actually shipped.

Four projects that prove the pitch. Open-source, regulated, responsible, and architectural. One of each.

01 · Shipping product

NextDialog

A calm interface for AI coding agents.

Multi-session desktop terminal manager for Claude Code, Cursor Agent, Gemini CLI, and custom agents. Built, shipped, and live at nextdialog.io.

Version
v0.3.2
Releases
17
License
MIT
Platforms
macOS · Linux · Windows
02 · Regulated-industry fluency

CareHub

HIPAA-compliant behavioral health SaaS.

A multi-tenant platform for behavioral health providers: clinical workflows, risk flagging, and audit trails that meet the rules without getting in the way of care.

Compliance
HIPAA
Domain
Behavioral Health
Scope
Clinical + Admin
03 · Responsible AI at real stakes

AskLisa AI

Constitutional AI with human oversight for Digital Citizen Academy.

An AI tutoring and guidance system built for an education partner, with rules, escalation paths, and human review loops that keep the humans in charge.

Model
Constitutional AI
Oversight
Human-in-loop
Sector
Education
04 · Graph knowledge architectures

Substrate

A personal cognitive data substrate.

A graph-based knowledge management system with spreading activation, designed to hold and connect personal and project context across time.

Architecture
Graph-native
Algorithm
Spreading activation
Scope
Cognitive infrastructure

How we work

Five phases. Hard gates.

Every engagement runs the same phase-gated discipline: evals before build, transfer at the end, and a written exit condition between every phase. No gate, no next phase. Fixed fees where it matters.

  1. 00

    Validation

    1–2 weeks

    Before anything gets built, the stated need gets reconciled against reality: the actual data, the actual systems, the actual people who'll live with the answer. Most AI failures are decided here, before a line of code exists.

    Gate: the brief survives contact with your data and systems

  2. 01

    Eval Definition

    1–2 weeks

    Passing and failing behavior gets defined as a version-controlled eval suite before the build begins. The suite is the spec. It's how a non-deterministic system becomes something you can accept, regress, and ship.

    Gate: no eval suite, no build

  3. 02

    Build

    3–6 weeks

    The system gets built against the suite: shipped into real hands early, instrumented from day one, iterated until the evals pass at threshold. Guardrails are operational, not theoretical.

    Gate: the eval suite passes at threshold

  4. 03

    Production Hardening

    2–3 weeks

    Monitoring live, failure modes rehearsed, integration load-tested, ownership assigned by name. The difference between a pilot and a system is everything in this phase.

    Gate: monitoring live, ownership assigned

  5. 04

    Transfer

    1–2 weeks

    You keep the eval suite, the architecture, the runbooks, and the working knowledge. The engagement is architected from day one to end, with your team running it without me.

    Gate: your team operates and extends it without me

Fit

Who this is (and isn't) for.

Good fit
  • Teams where AI has to be rigorous, auditable, and deployable the first time.
  • Organizations with committed AI spend that hasn't converted into shipped, measured value.
  • Healthcare, behavioral health, education, and other regulated or high-stakes domains.
  • Founders who need a working prototype in weeks, not a decision framework in months.
Not a fit
  • Staff augmentation or body-shop engagements.
  • AI readiness assessments, maturity models, or 20-page procurement roadmaps.
  • Projects that need a sales team, a marketing team, or a subcontracting layer.
  • Teams that haven't decided whether AI is worth doing. That's a different conversation.

82W takes three to four projects a year. Fit matters more than pipeline. If it's not a match, you'll hear that on the first call.

Start a conversation

Got a problem worth
shipping an answer to?

One email. Describe the problem, the stakes, and the timeline. We reply to everything that isn't a sales pitch.

Start a conversationbrian@82wdigital.com