CB / SOVEREIGN TERMINAL

Program delivery / Workflow design / Safe AI adoption

Christos Birkitt

Program, Workflow & AI Enablement Leader

I design auditable AI workflows, evaluation rubrics, and cross-functional delivery programs for Legal, Insurance, Finance, and Public Sector operations—helping organizations ship and adopt AI safely.

FOR HIRING MANAGERS, PMO LEADS & CONSULTING PARTNERS

Executive Delivery & 3-Page Governance Brief

Read delivery log Download executive brief PDF ↓

FOR TECHNICAL, LEGAL-TECH & EVALUATION DEEP DIVES

The Reality Ledger Protocol & Completed Legal Pilot RL-5P-v0.2

Inspect before & after Interactive Release Workbench HTML ↗ Method PDF ↓ Record JSON ↓ Pilot DOCX ↓

MODULE 01 / DELIVERY

Enterprise Delivery Log

Workflow design, data operations, and cross-functional execution across legal practice, nonprofit programs, federal assets, and international administration.

  1. Legal / AI enablement2023–PresentIndependent advisory engagement

    Plaintiff’s Litigation Practice (Confidential Client)

    Litigation Operations & Legal AI Enablement Consultant

    Managed civil litigation intake, motion and evidentiary record organization, and case-review workflows alongside trial counsel, identifying high-friction bottlenecks in manual medical and deposition summarization. Introduced structured two-stage LLM prompt templates and a human-in-the-loop verification checklist, modeled on the Reality Ledger rubric, to audit AI-generated case summaries for dropped qualifications and attribution drift. Reduced attorney verification rework by approximately 35% across sampled review packets while training legal staff on safe AI adoption and FRCP Rule 11 inquiry guardrails.

  2. Nonprofit / Data2017–2023Governance Policy Advisory: 2026

    Forma FGC (NFP)

    Project Manager & Data / Technology Consultant

    Personally designed and maintained SQL and Python workflows compiling raw geospatial and organizational data into structured geodatabases; managed website technology operations; and lectured alongside the FBI and DHS on data-protection protocols for at-risk populations affected by female genital cutting (FGM). Authored Forma FGC’s institutional AI and Data Ethics Governance Policy using NIST AI RMF 1.0 and ISO/IEC 42001 as alignment references.

  3. Federal / Operations2018–2019U.S. Secret Clearance

    U.S. Department of State

    Bureau of Overseas Buildings Operations (OBO) — Program Specialist / Realty Specialist

    Coordinated cross-functional workflows across diplomatic, legal, and technical stakeholders for international real estate acquisitions, facility planning, and geospatial/site risk analysis across high-security federal assets.

  4. Higher educationNew York / Athens

    New York University

    A.S. Onassis Program in Hellenic Studies & NYU in Athens — Program Administrator

    Redesigned administrative and student-intake workflows and managed budgets, vendor negotiations, and cross-border operations across New York and Athens to meet enrollment targets that increased by 100 percent.

  5. Property / Digital ops2016–2018

    Anderson Group Associates / Century 21

    Property & Digital Operations Manager / Licensed Realtor

    Managed commercial and residential transaction workflows, quantitative market valuations, client databases, and contract coordination alongside legal counsel.

  6. Courts / Civil litigation2015

    N.J. Superior Court, Hudson County Civil Division

    Judicial Intern — Chambers of Hon. Mary K. Costello

    Reviewed, organized, and summarized civil litigation motions, evidentiary records, and legal memoranda to support judicial rulings.

MODULE 02 / EVIDENCE

The Reality Ledger

A 60-second view of an operational circuit breaker: preserve the conditions, check the evidence, and re-scope the action so work can proceed under review.

Before / After

System case: RL-ILL-001, illustrative. Legal evidence: a constructed intake sequence and the completed exploratory pilot RL-5P-v0.2.

Panel A / Pre-repair

Unconstrained Workflow
Unguarded output

RL-ILL-001 / System-level caseIllustrative site-acquisition workflow

  • D-01Unconditional site-acquisition memo strips a 60-day provisional utility easement expiry.
  • D-02/03Retrieves a superseded 2022 flood report over the controlling 2025 High Flood Risk addendum.
  • D-04Asserts “no active litigation” despite a 0-character OCR drop on an 8-page docket.
  • REVIEWRecords a 14-second approval with no exception checks.

EIS 32.5

DISPOSITION: BLOCKED

Gates failed: Provenance / Coverage / Review / Profile

Legal intake / Constructed sequenceHow uncertainty hardens into an unsupported chronology

  1. STAGE 1 / WITNESS TESTIMONY

    “I remember water near Bay 3 around 11:30, but I cannot say whether that was Tuesday or Wednesday”

  2. STAGE 2 / COMPRESSION

    The day uncertainty is lost as the testimony is compressed.

  3. STAGE 3 / UNESTABLISHED CAUSAL SEQUENCE

    “Staff knew about water near Bay 3 by 11:30 on Tuesday, before the fall”

Panel B / Post-repair

Reality-Ledger Governed Workflow
Audited output

RL-ILL-001 / System-level caseSame illustrative case, repaired evidence and restricted action

  • D-04Re-OCR recovers the docket so the litigation claim can be checked against the record.
  • D-03Retrieval filter restores the controlling 2025 High Flood Risk addendum.
  • D-01The provisional expiry is preserved in Conditional_Caveats.
  • ACTIONUnconditional approval becomes a restricted informational diligence memo.

EIS 75.0

DISPOSITION: SUPERVISED CANDIDATE

Scorer & workbench verification: 3,378 deterministic JS-to-Python routing parity checks + 166 unit tests (139 Python, 27 JS). Controlled 16-packet synthetic benchmark (RL-BENCH-0.1): 8/8 fault recall (0/8 false positives) vs. 3/8 for retrieved-context-only Baseline A and 0/8 withheld for mean-only Baseline B.

RL-5P-v0.2 / Completed exploratory legal pilotManual evaluation / 5 passages / 10 AI summaries

Evaluated neutral and counter-framed summaries across five public civil opinions: Langford, Mendez, Thaqi, Cruz, and Pickens. Guardrailed prompts preserved uncertainty while manual review caught drift beyond binary checks.

4/4
Speaker doubt preserved
4/4
Cross-record dispute preserved
0/5
Unsupported relational inferences under guardrailed prompts
8/10
Outputs with subtle attribution, scope, or pronoun drift caught by manual review

Exploratory findings from this sample; not a generalized legal-performance estimate.

The system repair is an illustrative demonstration; the intake sequence is constructed; the five-passage legal pilot is a completed manual evaluation. Scorer & workbench verification: 3,378 deterministic JS-to-Python routing parity checks + 166 unit tests (139 Python, 27 JS). Controlled 16-packet synthetic benchmark (RL-BENCH-0.1): 8/8 fault recall (0/8 false positives) vs. 3/8 for retrieved-context-only Baseline A and 0/8 withheld for mean-only Baseline B. These controlled synthetic software results are separate from the pilot’s 10 manually evaluated legal summaries; post-adjudication routing uses supplied synthetic ratings.

MODULE 03 / GOVERNANCE

Auditable velocity.
Human-in-the-loop circuit breakers.

How do organizations protect reviewers from signing unverified machine summaries—and account for the operational time real review requires?

Working manuscript

The Right to Hesitate:
AI, Art, and the Work of Judgment

The practical governance problem is the handoff from a fluent machine summary to a human decision. A useful circuit breaker makes missing evidence, dropped qualifications, and unresolved disputes visible before a reviewer signs.

The operational aim is auditable velocity: define what can proceed, what needs verification, and who owns the exception. A restricted informational memo can keep diligence moving while an unconditional approval remains blocked.

REFERENCE MAP / From governance principles to review controls

Standards and legal references alongside their scope and a practical workflow interpretation. Reference links open the primary sources.
ReferenceReference scopePractical workflow interpretation
NIST AI RMF 1.0
MAP 2.3		 Scientific integrity and test, evaluation, verification, and validation (TEVV), including study design, data suitability, and construct validation. Define the rubric, evidence set, uncertainty categories, and evaluation limits before scoring.
NIST AI RMF 1.0
MEASURE 2.1, 2.3, 2.5		 Document test sets, metrics, and TEVV tools (2.1); evaluate under conditions resembling deployment (2.3); establish validity and reliability and document generalization limits (2.5). Retain source passages, prompts, outputs, rubric decisions, and scorer versions. Evaluate against review tasks and keep the exploratory pilot distinct from scorer verification.
NIST AI 600-1
MS-2.5-001		 Suggested action: avoid projecting capabilities from narrow, unsystematic, or anecdotal assessments. Report the five-passage pilot as exploratory. Preserve its sample size and denominators rather than treating it as broad validation.
NIST AI 600-1
MS-2.5-003		 Suggested action: check generated sources and citations during pre-deployment assessment and ongoing monitoring. Inspect source support, controlling versions, and missing records before accepting a summary’s claims.
NIST AI 600-1
MS-2.5-005		 Suggested action: verify training and TEVV data provenance and the grounding of fine-tuning or retrieval data. Track the provenance and versions of evaluation and retrieval artifacts. Artifact records alone do not establish coverage of training-data provenance.
ISO/IEC 42001:2023 An organizational AI management system (AIMS): establishment, implementation, maintenance, and continual improvement. Connect workflow controls to institutional policies, accountable owners, risk treatment, and continual improvement.
FRCP Rule 11(b) Presenting papers to a federal court entails certifications following reasonable inquiry, including legal and factual support. Preserve source checks, unresolved qualifications, and accountable attorney review before material enters court papers.

These references inform workflow design; this mapping is an implementation interpretation, not a certification or determination of legal compliance. The Waiting Room is the manuscript’s operational design concern, not a review-time method prescribed by these provisions.

Put auditable AI delivery into practice.

Discuss a program, a review workflow, or a safe AI adoption brief.

PROGRAM / WORKFLOW / AI ENABLEMENT