Full Technical + Research Overview
The tool that didn't exist.
Justice-impacted people face a hiring system that treats their record as a disqualifier -- not a data point. No tool combined AI career analysis, record-aware legal navigation, and disclosure coaching in one place. Until now.
This page is the complete documentation of what this platform is, how it works, what research grounds it, and how every decision was made. Open source. Nothing hidden. Designed to be read by scholars, funders, developers, and partner organizations.
The gap
70M+
Americans with a criminal record
<50%
Employed in year one after release
$7,500
Median first-year earnings post-release
0
Tools that combined AI career analysis, legal navigation, and disclosure coaching in one place -- before this
The full journey
Six stages. One platform. Every stage builds on the last. Free at every step.
The Forge
Upload a resume or describe your work history. Answer a few questions about goals, barriers, readiness, and record. The Forge outputs: a resume, a cover letter, career paths matched to your background, and a resource plan -- all in plain language.
Under the hood: Claude Sonnet (Anthropic) with a custom prompt scaffold that enforces the 10 behavioral rules. Resume parsed via OCR + NLP. Output hashed and logged.
Profile and Job Board
Fill in your contact info. Browse the verified fair-chance employer board -- real companies, real listings from several job sources.
Under the hood: Employers sourced from Airtable (SMR Employers base), imported to Neon PostgreSQL. Fair-chance verification is a manual human step -- not automated.
Application Tailor
Build job-targeted resume versions for specific listings. The AI tailors language to the job description while preserving your story.
Under the hood: Refinery Artifact system: each resume is versioned, stored with metadata (target job, company, application ID), and linked to the job application tracker.
Disclosure Planner
Plan when and how to talk about your record with specific employers. The planner walks through timing, framing, and what to say -- grounded in actual Ban-the-Box law and fair-chance employer research.
Under the hood: Structured coaching dialogue with t.ROY. Output saved as a disclosure_plan artifact. Nothing in the practice transcript is stored.
Interview Practice
Text or live voice mock interviews tailored to your target role. Behavioral, situational, and disclosure-specific questions. Real-time feedback on framing, specificity, and narrative arc.
Under the hood: Text mode: Claude Sonnet. Voice mode: OpenAI Whisper (transcription) + Claude (feedback). Practice audio is never stored -- only the frame used and whether the point landed.
Applications + Follow-Up
Track every application: company, status, dates, which resume version you used. Generate follow-up emails and save them to My Materials.
Under the hood: job_application table in Neon. Refinery Artifact system for follow-up emails and cover letters. My Materials = the vault of all your generated documents.
The behavioral protocol
t.ROY follows 10 non-negotiable rules on every interaction. These are not personality guidelines -- they are hard constraints derived from peer-reviewed research and Troy's decade of direct service. They cannot be overridden by user prompting.
Name the feeling before solving the problem
Affect labeling reduces amygdala reactivity before high-stakes work begins (Lieberman et al., 2007).
Strengths before gaps -- always
Deficit-first framing activates threat response. Asset-first activates approach motivation.
Never volunteer legal interpretations
t.ROY is a career guide, not a lawyer. It flags disclosure moments and points to resources.
Match complexity to the user's stated readiness
Cognitive load theory: overwhelming someone mid-crisis loses them permanently.
Offer the next step -- never a lecture
Autonomy-supportive language predicts sustained engagement (Deci & Ryan, 2000).
Use plain language. Always.
The platform is written at a 6th-grade reading level. People in crisis don't need vocabulary tests.
Redemption arc, not deficit report
People who construct bad-to-good story arcs show measurably better post-release outcomes (McAdams, 2013).
Consent before every data use
Surveillance trauma is real. Trust must be built, not assumed.
Every recommendation requires a reason
Unexplained AI advice erodes trust. t.ROY explains its reasoning in plain language every time.
If someone is in crisis, stop and redirect
A crisis disclosure during career coaching requires a human, not a chatbot. Full stop.
What makes it different
AI career analysis built for justice-impacted people -- employment gaps, records, and barriers included in the model
vs. Generic resume builders that treat a 10-year gap as a red flag
Record-aware disclosure coaching baked into the Refinery -- not an afterthought
vs. Job boards that surface Ban-the-Box listings but give no guidance on what to say when asked
Every AI recommendation logged with input hash, model ID, explanation, and latency -- full audit trail
vs. Black-box recommendations with no accountability
10 research-grounded behavioral rules that govern every t.ROY response -- non-negotiable, documented
vs. A generic chatbot tuned for engagement, not outcomes
Consent-gated partner dashboard -- organizations can only see clients who have explicitly opted in
vs. Program tools that harvest participant data without meaningful consent
Free at every step -- no paywall, no credit card, no data sold -- ever
vs. Freemium tools that paywall the parts people actually need
Built by someone who navigated the same system -- not designed at a distance
vs. Tools designed by people who have never been incarcerated, for people who have
The privacy model
Surveillance trauma is real. Many justice-impacted people have been monitored, recorded, and reported on without their knowledge or consent for years. This platform is designed to actively undo that pattern.
Practice tools store frames, not words
The Disclosure Planner and Interview Practice tools store whether a coaching frame was used and whether the key point landed -- never the user's actual answers, transcript, or audio. The system can show you progress without survelling your process.
Partner access is consent-gated
Partner organizations (nonprofits, reentry programs, job centers) can only see clients who have explicitly opted in to share their progress. The toggle is in Settings. Off by default. Cannot be enabled by the partner -- only by the user.
Forge data stays local until you choose otherwise
The Forge runs client-side. Resume data, goals, and story answers are stored in your browser's localStorage -- not our servers -- until you create an account and choose to sync. You can complete the Forge and never create an account.
No data sold -- ever
Steel Man Resumes LLC is not funded by advertising. There is no data monetization layer. This is documented in the AGPL-3.0 license and the source code is public.
Right to deletion
Settings > Privacy > Delete all data removes everything server-side and clears localStorage. The deletion is immediate and irreversible.
The research foundation
Every design decision is grounded in peer-reviewed research. The behavioral protocol, the copy reading level, the narrative arc of the Forge output, and the privacy model all trace to specific empirical literature.
Affect labeling reduces amygdala reactivity by up to 50%
Lieberman et al., 2007 -- UCLA
Putting feelings into words activates prefrontal cortex and dampens the stress response. The Forge uses narrative reframing at every step -- not as therapy, but as a documented performance technique that lowers cognitive resistance before high-stakes work.
Redemption narrative sequences predict higher well-being and post-release stability
McAdams, 2013 -- Northwestern
People who construct bad-to-good story arcs show measurably better outcomes across health, relationships, and stability. The Forge output is structured as a redemption arc, not a deficit report. This isn't a design flourish -- it's a replication of a documented psychological mechanism.
Autonomy, competence, and relatedness are the three non-negotiable psychological needs
Deci & Ryan, 2000 -- Self-Determination Theory
Incarceration systematically strips all three. This platform rebuilds them: user-driven choices (autonomy), skills surfaced from real history (competence), and language grounded in lived experience rather than clinical distance (relatedness).
Stigma-based discrimination persists even in Ban-the-Box jurisdictions
Agan & Starr, 2018 -- American Economic Review
Removing the checkbox doesn't remove the bias. The Refinery's disclosure coaching prepares users for the conversation that happens after the application -- not just the form.
Cognitive load undermines decision quality under economic stress
Mullainathan & Shafir, 2013 -- Scarcity
Scarcity (of money, time, housing stability) taxes cognitive bandwidth directly. The platform is designed for people under maximum load: short prompts, no jargon, one question at a time, no decision fatigue.
Workforce reentry programs with wraparound support show 2-3x better employment retention
MDRC / CEP Meta-Analysis
Employment alone doesn't stick. The Refinery is a wraparound: resume + disclosure + interview + applications + coaching, in one place, free.
Architecture and accountability
Every AI recommendation is logged with an input hash, model ID, explanation, and latency. Full audit trail. No black box. Built for funder reporting, compliance review, and program evaluation.
// Every AI decision logged:
decision_log {
input_hash: sha256(user_input)[0:16]
model_id: "claude-sonnet-4-6"
explanation: "why this recommendation"
latency_ms: 1247
user_action: null | "accepted" | "modified"
}
Neon PostgreSQL -- all structured data (users, profiles, artifacts, applications, consent records, decision log)
Cloudflare R2 -- encrypted document storage (uploaded resumes, generated PDFs)
Vercel -- serverless hosting, edge functions, preview deployments
Resend -- transactional email (magic links only -- no marketing email)
Airtable -- verified employer data source (manually curated, imported to PostgreSQL)
For partner organizations
Reentry programs, workforce development orgs, legal aid societies, and job centers can integrate Steel Man Resumes into their client workflow. Access codes unlock higher AI usage limits and allow consent-gated progress tracking.
Access code
One code per program. Clients enter it at sign-in to unlock higher limits. Codes are linked to your partner account.
Partner dashboard
See cohort progress -- stage, artifacts completed, last activity -- for clients who have opted in to share. CSV export for program reporting.
No vendor lock-in
AGPL-3.0 license. The source code is public. You can self-host a private instance for your program.
To get an access code
Email Steel Man Resumes LLC directly. Free for nonprofits and community organizations during the launch period.
For funders
Steel Man Resumes LLC (WI) is currently transitioning to a nonprofit structure. The platform is AGPL-3.0 open source and free to all users. No revenue is generated from user data.
Measurable outputs
Every Forge completion, resume built, disclosure plan, interview session, and application tracked -- with user consent and without collecting PII in aggregate reports.
Audit trail
The decision_log table captures every AI recommendation with full provenance. Funders can request reports on AI performance, model usage, and outcome signals.
Target population
Justice-impacted adults in Wisconsin (initial), Midwest (expansion), national (AGPL self-host). Primary focus: people within 90 days of release or recently released.
No paywall -- by design
The platform will never charge users. Sustainability comes from organizational licensing, grants, and eventual nonprofit status.
Built by someone who lived it
Steel Man Resumes was built by Troy Carr -- a justice-impacted workforce advocate with a decade of direct service navigating the same system these tools are designed to help people survive.
The platform reflects what he wished existed when he was in it: one place that understood records, didn't treat gaps as disqualifiers, and gave plain-language guidance instead of generic advice designed for people who have never been incarcerated.
The research is real. The behavioral rules are real. The employers are verified. The privacy model was designed with surveillance trauma in mind because Troy has lived that too.
Steel Man Resumes LLC -- Wisconsin -- AGPL-3.0 open source