Statement of Prior Experience

Entity of record

USA Govvy Inc. is the name of record. Every engagement below was performed by the same principals and the same delivery team, under the entity's prior operating name where the commit record predates the current filing.

The record covers December 2025 through August 2026. It comprises five engagements, sixteen distinct systems, approximately 280,000 lines of application code, and more than 720 automated tests across 285 commits.

Every system in the record runs standalone. Local inference against the customer's own model weights, local embeddings, a self-hosted data plane, and default-deny egress — no outbound call to a commercial endpoint. Section 3 sets out how, and why the architecture makes that a configuration change rather than a port.

Every count in this document was taken from the filesystem, not from a marketing page.

INHERIT — Federal Research IP Discovery Platform

The problem

Every year the federal government invests $60 billion into university research. Universities return $3 billion in licensing revenue. That is five cents on the dollar. Over 95% of federally funded inventions are never commercialized.

This is not a market failure. It is a stewardship failure, and it has three causes.

The stewards are overwhelmed. Major universities receive 500 to 3,000 invention disclosures a year. The average technology transfer office with a budget under $1M has 2.4 staff. Licensing headcount fell 4% in 2024 while disclosures rose.

The knowledge is scattered. The technologies live in fragmented databases across NIH RePORTER, NSF Awards, USPTO, SBIR.gov, DOE labs, and hundreds of individual technology transfer sites. No single search covers them.

Evaluation cannot scale. Assessing one asset takes 20 to 40 hours of manual work — patent claims, science review, IP status, principal investigator, office posture, market size, competitive landscape. Professional IP scouts charge $500 to $2,000 an hour.

The solution

The five-minute path from federal research IP to drafted founder outreach.

INHERIT crawls federal research archives, scores assets across six dimensions, scans the live application market for products each asset could displace or extend, and lets an operator converse with an agent grounded in whatever asset, category, or market gap is in front of them — producing decks, one-pagers, whitepapers, and outreach drafts without leaving the canvas.

Capabilities delivered

Fifty-one purpose-built federal and international data connectors. These are working async API clients, wired and operating.

Platform surface. 24 API routers exposing 96 route handlers. 41 database tables. 22 SQL migrations. 21 CLI tools. Nine agent subsystems — canvas, discover, enrich, evaluate, ground, market, measure, remix, steward.

Operator canvas. Eleven slash commands across seven anchor types. /deck builds a ten-slide pitch deck; /onepager; /whitepaper; /memo; /critique scores against a five-dimension rubric; /email pi|customer|investor; /sharpen; /research; /landscape; /brief; /why.

Retrieval architecture. Ten enrichment modules — citation crawler, DOI enricher, entity resolver, full-text extractor, golden record builder, inventor enricher, URL enricher, market intelligence, attention tracker. Sentence-transformer embeddings at 384 dimensions with cosine similarity retrieval. Fuzzy deduplication and reranking. A three-layer memory architecture: sliding conversation window with compression, re-loaded anchor context, and a persistent user profile.

Provenance. Field-level lineage on every extracted value — source, SHA-256 value hash, confidence score, extraction method, raw evidence, timestamp. Citation edges and source freshness tracked as first-class tables. Source URLs revalidated with tombstoning.

Security. Passwordless magic-link identity: no passwords, single-use tokens with a fifteen-minute TTL, thirty-day HttpOnly SameSite sessions. Same-origin proxy architecture keeping the session cookie first-party by design. A credential scrubber running on every inbound message before it touches a database row, a log line, or the prompt stream — detecting provider keys, AWS access keys, Bearer tokens, password assignments, and Luhn-valid card numbers, and replacing them with typed redaction markers. Per-plan rate limiting on rolling twenty-four-hour windows. Incognito mode bypassing persistence entirely.

Continuous integration. Least-privilege workflow permissions, read-only by declaration. No API keys in CI on purpose, so model-backed paths must exercise their deterministic fallbacks. An anti-false-green gate that fails the build if a test silently skipped. A separate Postgres portability job running against a real postgres:18 container, added after a production incident, with the postmortem written into the workflow file.

Stack. Python 3.11, FastAPI, SQLAlchemy 2.0 async, Pydantic 2, Typer, WeasyPrint. Next.js 15 App Router frontend. Neon Postgres 18. Sentry observability with releases pinned to commit. Stripe billing. Render infrastructure.

Relevance to federal work

This is the strongest federally citable asset in the portfolio and it is not close. The system already reads DoD, DARPA, NASA, DOE, NIH, NSF, SBIR, USAspending, USPTO, NIST, and FDA archives in production. It handles the ingestion, normalization, deduplication, entity resolution, and scoring of federal records at six-figure scale. A contracting officer evaluating whether this team can work with government data does not have to take our word for it.

It also does something rare. It ships a health endpoint that publicly reports its own defects — ingest freshness, coverage gaps, data anomalies — rather than presenting a green light. That is the operating posture a federal customer should want in a vendor.

Maven IT Agentix — Isolation-First Clinical AI Platform

The problem

Clinicians spend close to two hours on documentation and computer work for every hour of direct care. It is a leading cause of burnout. Administrative work is roughly a trillion dollars a year in U.S. healthcare spend.

The common way to deploy clinical AI has three flaws. Most tools share one runtime. Everyone's work flows through one pool of memory and compute.

It accumulates state. The more it is used, the more working data builds up, and the system slows and costs more. It blurs boundaries. When all work shares one space, keeping one patient's data from reaching another's becomes a constant fight. And it fails wide. One bad task or one overloaded component degrades service for everyone.

Under HIPAA, "we try to keep data separated" is not a sufficient control.

The solution

A clinical AI platform where every task runs in its own isolated, ephemeral container.

Maven gives each clinical task a private agent. The agent is created on demand. It loads only the data it needs. It does the work. A clinician approves the result. Then it is destroyed. Nothing carries over.

The lifecycle is fixed: PROVISION → HYDRATE → EXECUTE → PERSIST → TEARDOWN → SETTLE.

The system does not clog, because working memory never outlives its workspace. Patient data cannot cross over, because isolation is enforced at two independent layers — the database credentials a task holds, and row-level security inside the database itself. A cross-patient leak would require both layers to fail at once.

Capabilities delivered

Five architectural layers. Experience (Next.js 15 / React 19 clinician console), Gateway, Orchestration, Isolation, and Memory & Integration. Nine workspace packages — agent-runtime, approval, model client, logger, manifest, MCP servers, memory-client, phi-guard, types.

Tiered memory subsystem. Working memory lives in the sandbox and dies with it. Episodic memory is scoped per patient and provider. Semantic memory is scoped per provider. Retrieval is scoped hydration — load only task-relevant memory, never load the world — with sub-500ms hydration proven by integration test.

• ENABLE ROW LEVEL SECURITY and FORCE ROW LEVEL SECURITY on every PHI table
• Policies keyed on (tenant_id, agent_id) read from LOCAL session GUCs with missing_ok = true, so an unset scope matches nothing — default-deny by construction
• The application connects as a role created NOSUPERUSER NOBYPASSRLS
• PHI columns stored as encrypted bytea in AES-256-GCM iv||tag||ciphertext layout
• An explicit append-only, PHI-free memory access log
• Fifteen integration tests running against a live pgvector container in continuous integration

Ten security invariants, enforced or explicitly tracked. No mutable image tags — digest-pinned, with a red-team test. Explicit egress allowlist, default-deny. No PHI in logs, enforced by a dedicated scanner. No PHI in browser storage, enforced by lint rule and scanner. Row-level security on every PHI table. Default-deny authorization. No any or @ts-ignore. Supervisor exposing exactly three methods. Signed images only. mTLS between services. Platform code capability-agnostic, enforced by scanner.

Single-path model rule. Every inference call routes through one package. Direct SDK calls elsewhere fail review, enforced by an invariant test. Prompt caching is mandated on the stable system block, with per-tenant cost telemetry.

Human in the loop by construction. No AI output reaches a patient record without explicit clinician approval. The approval gate is socket-independent. Edit statistics are captured PHI-free.

Continuous integration. Three merge-gating jobs: build, typecheck, lint, unit tests, security invariants, and red-team suite; an integration job against a live pgvector/pgvector:pg16 container that creates the NOBYPASSRLS role and applies migrations before running the memory suite; and a supply-chain job for image signing and SBOM.

Stack. TypeScript strict, Node 22, pnpm and Turborepo, Next.js 15, React 19, PostgreSQL 16 with pgvector, model routing with escalation tiers, embeddings at 1024 dimensions, Vitest, Model Context Protocol.

Relevance to compliance work

This is the closest thing in the record to the evidence-layer thesis written as engineering. Three transfers are direct.

Access control maps onto NIST SP 800-171 AC. Forced row-level security with a non-bypassing application role and default-deny scoping is not a policy statement. It is a database that refuses.

Audit maps onto AU. An append-only access log that records the fact of access without recording the sensitive content is the pattern an assessor wants to see.

Encryption maps onto SC. AES-256-GCM at rest with tenant-scoped envelope keys and TLS 1.3 in transit, with PHI barred from logs, URLs, browser storage, and analytics by scanner-enforced rule.

And the platform publishes its own enforcement audit — which invariants are wired, which are deferred, and by what phase. A vendor that grades itself honestly in its own repository is a vendor whose self-assessments a government customer can price.

Clarity MedLegal Partners / Clarity AI

The problem

The medical-legal market spends $15.6 billion a year on manual work. No platform covers the full workflow.

A QME practice doing forty evaluations a month has forty reports to audit, forty billing packets to assemble, forty checklists to track, and a hundred and twenty deadlines to monitor. A personal injury practice managing two hundred active patients has two hundred authorization windows, two hundred treatment plans, and a billing pipeline that leaks revenue every day something slips.

The doctor is in the exam room. The attorney is in court. The paralegal is on the phone. The biller is buried in CMS-1500 forms. Nobody has time to check whether the supplemental report was audited, whether the ML-205 packet was submitted, or whether an authorization expires next Tuesday.

The solution — four systems, one lifecycle

A. Clarity SI — Med-Legal Intelligence System

February 2026 · 57 files · 5,302 lines · counts verified against the filesystem

A domain intelligence system installed in one command, giving an operator fifteen commands. Each does something that currently takes hours or costs thousands of dollars.

Type /chronologize. It reads five thousand pages of medical records and builds an interpretive timeline. Type /demand. It writes a jurisdiction-aware demand letter anchored to comparable verdicts. Type /pipeline. It runs every module in sequence and hands back a complete case intelligence package.

Twelve skills: adversarial strategy, billing forensics, case merit intelligence, causation nexus engine, compliance sentinel, demand generation engine, expert witness network, IME/QME command, lien resolution and MSA, life care projection, medical literature standard of care, medical record intelligence. Fifteen commands. Eight orchestration agents. Six marketplace connectors. One orchestration engine.

The eighth agent is the Compliance Sentinel. It monitors everything, cannot be turned off, and holds override authority. It runs all the time, can halt any other agent, keeps the audit trail, and checks every data handoff.

A deliberate architectural choice sits at the center. A personal injury case can carry five thousand pages of medical records, and the system reads the entire file at once. No chunking, no lost cross-references. It finds the note from Provider A that contradicts the report from Provider B.

B. Clarity SI Gateway — Persistent Operations Daemon

February 2026 · 105 files · 5,703 Python LOC · 57 automated tests

A persistent Python daemon connecting the intelligence layer to the real world. Doctors, attorneys, paralegals, billers, and office managers send commands in plain English over WhatsApp and Slack. Results come back as structured messages and branded PDFs.

• Receives messages from WhatsApp via Twilio and from Slack via the Bolt SDK
• Understands intent through two-stage NLP — a regex fast path with a model fallback
• Executes commands by injecting domain skills into inference calls
• Generates billing packets, medical reports, chronologies, and demand letters as branded PDFs
• Follows up proactively with deadline alerts, authorization renewals, case health scores, and next-step suggestions

Twenty-six verified commands across two practice lines. QME workflow: intake, audit, bill, supplement. PI workflow: intake, bill, report, authorization. Cross-cutting: chronologize, demand, audit-bills, find-expert, causation, life-plan, merit-screen, research-soc, resolve-liens, pipeline, war-room, adversarial, compliance, conquest, checklist, day-sheet, recommend. Each carries natural-language keyword mapping.

Nine containerized services — gateway, four workers (commands, pipeline, documents, messaging), scheduler, Postgres, Redis, object store. Ten data models. Five worker queues.

Security implementation. AES-256-GCM encryption with PBKDF2-HMAC-SHA256 key derivation at 600,000 iterations. TLS 1.3 on every endpoint. PHI detection with automatic redaction in logs. Role-based access control across three roles and nine permissions, enforced on every endpoint, with the full matrix published. Append-only audit trail. Per-patient, per-case HIPAA authorization with e-signature, validated before every data pull. A credential vault holding a 32-byte master key in the OS keychain, with AES-256-GCM and unique 96-bit nonces, rotation supported, and a documented ninety-day rotation command. Production key material moves to AWS KMS, Azure Key Vault, GCP Cloud KMS, or HashiCorp Vault.

Proactive engine. Deadline monitoring every six hours with alerts at 90, 60, 30, 14, 7, 3, and 1 days across four urgency tiers, escalating to the attorney if a critical alert goes unacknowledged for twenty-four hours. Case health scoring every Monday at seven across five dimensions. Follow-up analysis after every command, emitting five flags: treatment gap, billing anomaly, incomplete note, MMI reached, authorization expiring. Daily firm digest at eight, weekly Monday roll-up.

C. Clarity MGT — Provider Management Platform

July 22 – August 6, 2026 · 86 files · 3,927 lines of research artifacts · 8,610-line prototype · 27 commits

A greenfield healthcare provider management platform replacing two legacy spreadsheets. Two modules, four roles.

Provider Scheduling — rolling twelve-month calendar, seven entry types, three-month lock rule, change requests, manager approval inbox, audit trail. Concerns & Activity — concern submission, medical-assistant auto-routing, four statuses, activity trail, provider summary, priorities. Roles — remote team, medical assistants, providers, one scheduling manager.

The method is the deliverable as much as the product. A six-phase information architecture sequence, one commit per phase, each run as a multi-agent swarm with adversarial verification, against a published definition of done:

Tree test clears ≥80% success on every critical task with real participants. Every top-level label has a verbatim behind it, cited from the Verbatim Ledger. Nothing sits deeper than three meaningful levels; breadth over depth. Never validated via visual prototype — tree test first.

Artifacts produced: content inventory with redundant / outdated / trivial flags and a verbatim ledger; findability diagnosis across four role lenses with eleven findings; open card sort protocol with thirty-one cards; three candidate architectures plus the incumbent, with a trigger-word audit and thirteen labeled assumptions; tree test protocol with eleven tasks, five critical, at an 80% gate; navigation specification; three critical task flows; heuristic evaluation by three blind evaluators against Nielsen's ten; and a gate review closing nine blockers. Then ten front-end prototype commits with WCAG and NN/g quality gates, a twenty-one-viewport responsive matrix, and a 120-render QA pass.

D. Clarity AI — Evidence Layer Research

August 2026

The thesis: simulating a procedure is now free; proving a human is competent to perform it is not. The product is the Procedural Competency Record — a signed, versioned, portable artifact stating that a named clinician, on a named date, on a named device, met a named standard.

Supporting it is an Evidence Ledger built on an explicit four-tag discipline: SOURCED, SOURCED-SECONDARY, ESTIMATE (inference or arithmetic, derivation shown, not a fact), and GAP (searched for, not found publicly, named as diligence work rather than filled in). Twenty-six sourced claims, seven named gaps, five tagged estimates. Every row carries the claim, its tag, its sources, a "do not misuse" field, and why it matters. It includes a self-authored FDA 510(k) clearance census with a stated limitations section acknowledging that De Novo grants are invisible to the data source, so every count is a floor.

Relevance to compliance and defense-adjacent work

Med-legal work is evidence work. The product is a record that survives examination by a hostile third party — every assertion traced to a source document, dated, versioned, and defensible under questioning. The client's own operating language is source-linked and defensible. That is structurally the same deliverable as an 800-171 traceability matrix: 110 requirements mapped to 320 assessment objectives mapped to named, dated artifacts, exported to an assessor on demand.

Regulated-boundary AI. The application runs against PHI under HIPAA. The constraint — build inside a compliant boundary with correct domain logic, because the data cannot enter general-purpose commercial tools — is the same constraint DFARS 252.204-7012 imposes on CUI.

Multi-vendor delivery governance. Standard operating procedures flowed down uniformly to subcontracted developers across three vendor organizations and two countries, on two-week sprints with backlog gating on acceptance criteria. That is the flowdown discipline a prime is scored on.

Claim-tagging as an operating habit. The Evidence Ledger's SOURCED / ESTIMATE / GAP discipline, with derivations shown and gaps named rather than filled, is the same epistemic posture a self-assessment under CMMC requires — and the same one most vendors cannot demonstrate.

Port San Antonio

The problem

Port San Antonio operates one of the largest innovation campuses in the United States — 1,900 acres, 80+ tenants, $5.6 billion in annual economic impact, tenants including the 16th Air Force, Boeing, Northrop Grumman, and Leidos. Its digital presence did not reflect the scale of the physical operation.

Four distinct audiences share one front door: tenants and businesses, investors and developers, students and workforce, and community. Tenant services were buried under real estate. Contracting and procurement had no action-oriented entry point. Job seekers and vendors had to reason about the org chart to find what they needed.

The solution — four delivered systems

A. Information architecture restructure across the full campus portal

40 pages · 13,419 LOC · three dated revisions

The six-item navigation was rebuilt into a six-item, thirteen-column, thirty-link audience-segmented mega-menu, with a descriptor line written for each audience.

The first revision elevated Campus to top level for campus-first positioning, created Work With Us as an action-oriented destination for employment and contracting, restructured Education with literal labels, moved Tenant Services out from under Real Estate into the Campus menu, added a Talent & Workforce column, and simplified About and News. Six new pages were created; eight items were pruned from navigation with files preserved and reachable by direct URL and footer.

The second revision split About into two columns — About Us and Leadership. The third added Kelly Field (SKF) Airport under Campus and collapsed a redundant Industries column that duplicated the Education dropdown.

B. Three complete alternate design directions

123 pages · 34,279 LOC · three independently deployable environments, all live

Each direction is a full forty-one-page copy of the site with its own design system — genuinely divergent, not restyled. Editing one direction's stylesheet restyles all forty pages in that direction consistently. Each deploys independently; pushing one never touches the others or the control. The typographic register shifts deliberately across directions, down to the casing of the hero line.

C. UX Visual Reference Guide

April 2026 · formal client-facing design research deliverable · live

Six sections. A curated reference library mapping the design language of ten award-winning sites to the Port's digital direction — built around the Port's own brand identity.

• Brand Foundation. An eight-token color system derived from the Port's swoosh logo. Brand red replaces the arbitrary gold-and-blue pairings common in port authority design and gives the digital experience a distinct personality.
• Primary Reference. One benchmark site analyzed across six design signals.
• Extracted Design Language. Six named principles, each with a Port-specific application. Darkness as premium signal. Typography as navigation. Geography as hero — 1,900 acres deserves cinematic framing. Filtering as experience design. Concierge language architecture. Scroll as storytelling.
• Reference Library. Ten annotated award-winning sites, each carrying award status, extracted pattern, and a Port application note.
• Synthesis. An eight-row prescriptive directive table across color, typography, motion, scroll architecture, discovery, content architecture, copy tone, and navigation.

Two directives bear directly on architecture: content order runs atmosphere → story → proof → action, with economic impact data appearing after emotional buy-in, not before; and user types — tenant, investor, visitor, job seeker — are wayfinding categories, so let users self-select before they read.

D. GTM Intelligence Console

July 2026 · 1,889 lines, single-file zero-dependency application · live

Website visitor capture, lead intelligence, and human-approved agentic outreach. Seven views: overview, visitors, leads, approvals, tours, agents, sequences. Four named agents, with a guarantee written into the product copy:

A seven-stage conversion funnel written in plain English: visited the site ("people we could not name") → matched to a company ("now we know who they are") → good fit ("score of 70 or higher") → opened or replied ("they wrote back") → booked a tour → toured the campus ("they showed up") → signed ("lease or letter of intent").

Six outreach sequences aligned to the campus's actual clusters — Aerospace Expansion, Cyber Cluster, Defense Expansion, Space & Test, Advanced Manufacturing under CHIPS, and Startup Landing. Five sector segments. Command palette with full ARIA combobox semantics. Guided walkthrough. Five-step onboarding dialog. Light and dark theming. Skip-to-content. Empty states written in plain language.

Relevance to public-sector work

Direct working exposure to the Port's audience structure, tenant mix, procurement pathway, and cybersecurity industry cluster. Contracting and procurement findability was a designed navigation objective — this team has already studied how vendors locate contracting information at this specific authority.

The method transfers on its own terms. Contracting officers evaluate whether a system can actually be used by the people who must use it. A documented task-success protocol with pass thresholds is the instrument that answers that question, and this team writes them.

Portwize — Agentic Operating System for Global Trade

The problem

Port data is fragmented, stale, and non-auditable. Three consequences follow.

Cost leakage — millions in preventable detention and demurrage fees. Operational chaos — wasted truck turns and inefficient berth allocation. Trust deficit — no evidence-grade data for insurance and trade finance disputes.

Drayage does not fail because people do not care. It fails because the information needed to act lives across portals, emails, PDFs, and spreadsheets, and nobody can see the whole clock.

The solution

A vertical operating system for port commerce. It ingests public and private event streams — AIS, terminal operating systems, gate events, rules. It resolves entities — vessels, voyages, containers — into a single timeline. It computes signal products with high freshness. And it delivers through a direct interface, an API, webhooks, and auditable evidence bundles.

We do not sell visibility. We sell signals that are predictive, actionable, and auditable.

Without a truth layer, AI is guessing. Portwize provides the operational certainty that high-value settlement requires.

Capabilities delivered

Thirty-four autonomous agents across three transport modes.

One hundred forty-six HTTP endpoints across fifty route modules — agents, alerts, API keys, auth, SSO, users, entities, events, evidence and evidence verification, geofences, health, metrics, packs, signals, security, integration, billing, compliance audit, data sovereignty by region, FinOps cost attribution, EDI and TMS integrations, ML ground truth and models, simulation scenarios, webhook subscriptions, and a WebSocket stream.

Seven API middleware layers — auth, CSRF, metrics, rate limiting, RBAC, security headers. Seventeen core sub-packages — alerts, audit, auth, billing, data sovereignty, entity resolver, evidence, FinOps, integrations, multi-tenant, packs, security, signals, simulation, validation, webhooks.

Enterprise integration adapters — Oracle Transportation Management, SAP Transportation Management, and EDI with X12 and EDIFACT mappers and a gateway. Each with adapter, client, mapper, and webhook handler.

Twenty MLOps modules — vessel ETA predictor, dwell time predictor, gate congestion predictor, noise model trainer; ground truth collection and matching; feature extraction and pipelines; drift detection, performance tracking, and alerting; retrain orchestration and model versioning.

Infrastructure as code — twelve Terraform files across seven modules: VPC, EKS, RDS, ElastiCache, MSK, S3, IAM. Plus Docker, Kubernetes manifests, and Prometheus.

Data providers wired. Maritime AIS across four providers. Aviation ADS-B through the OpenSky Network with OAuth. Weather through NOAA and OpenWeatherMap. Traffic through the HERE Platform with routing. Customs, terminal, truck, warehouse, and rail integration keys. Firmographic enrichment. Stripe billing. Prometheus monitoring.

Five named signal products — arrival window confidence, anchorage queue index, gate performance index, disruption detector, ready-for-pickup confidence.

Operator interfaces. A sixty-six-component console across sixteen route segments — dashboard, signals, signal matrix, shipments, shipment thread, operations, evidence, simulation, FinOps, billing, alerts, packs, partner portal, integration, API. A separate twelve-page maritime command console — operational command with signal maturity and confidence metrics, vessel tracker with ETA prediction and risk assessment, terminal intelligence with benchmarks and anomaly detection, signal catalog, alerts and webhooks, and an operations assistant.

A production marketing site — twelve pages, eighteen components, a dual-path onboarding wizard branching between customer and investor personas with a shared context provider, animated step transitions, progress indication, and a slide-to-confirm primitive.

A documented design system — six core color tokens with HSL custom properties, a 4px baseline grid with eight spacing tokens, a twelve-column responsive grid scaling to six and four, four button variants including an animated neural shimmer reserved for agent actions, four status badge states, three signature effects, two layout patterns, two page templates, and five named animations with fixed durations — all rendered as a live five-section reference application.

An agentic RAG backend — LangGraph workflow orchestration over ingestion, policy, and dispute agents; FAISS vector store; PDF processing pipeline with staged input, processed, and failed directories; Pydantic state models; Supabase persistence.

A ten-port US maritime intelligence system — an eighteen-field port data model covering vessel counts, waiting vessels, average wait hours, berth utilization, TEU by day, month and year to date, alerts, news, social mentions, and a data quality score. Twenty-six named sources across five source classes: ten official port authorities, six maritime trade publications, four vessel tracking services, four industry and government research bodies, and social channels. A seven-section analytics dashboard built without external chart libraries. And a 1,537-line analytical intelligence report covering ten major US ports.

Relevance to federal and public-sector work

Ports are critical infrastructure. This platform ingests public data sources — AIS, ADS-B via OpenSky, NOAA weather — and resolves them into an auditable timeline with evidence bundles designed to survive a dispute. That is the same architecture a federal customer needs for supply chain risk visibility, and it is already built against public-first data, which is the posture that clears an authorization boundary fastest.

The enterprise adapter work — Oracle OTM, SAP TM, X12 and EDIFACT — is direct experience integrating with the systems federal logistics contractors actually run.

Sovereign deployment — standalone operation on private models

Every system in this record is built to run inside a customer's boundary, on the customer's hardware, against the customer's own model weights, with no outbound call to a commercial inference endpoint.

This is an architectural property, not a port. It holds because each system was built to one rule.

One seam, by design

Each system routes every model call through exactly one adapter. There is no second path.

• INHERIT — a single SDK wrapper. All nine agent subsystems call through it. Nothing else touches a model.
• Maven IT Agentix — every call routes through one package, and the rule is enforced by an invariant test. A direct SDK call anywhere else fails the build, not review.
• Clarity SI Gateway — one client, with domain skills injected as system prompts at call time rather than compiled into the binary.
• Clarity SI — markdown only. Nothing to compile, nothing to configure, runtime-agnostic by construction.
• Portwize Signal — the core signal platform carries no model dependency at all. It is deterministic. Signals compute without inference.

Substituting a locally hosted model is a change to one adapter and one environment variable. It is not a rewrite, and it does not fan out across the codebase, because the codebase was never allowed to fan out.

Inference runs where the customer says it runs

The adapter targets an OpenAI-compatible HTTP interface, which is what every serious local runtime speaks — vLLM, llama.cpp, Ollama, Text Generation Inference. Open-weight models on customer GPUs. Weights on customer disk. The boundary holds because nothing crosses it.

Embeddings are already local. INHERIT runs sentence-transformers all-MiniLM-L6-v2 at 384 dimensions in process — no API call, no vendor, no egress. Maven's embedder is a port with a working adapter behind it. Neither system has ever required a hosted embedding service, so neither has to give one up.

Degraded operation is a tested path, not a discovery. INHERIT's CI runs with no API keys on purpose, forcing every model-backed path through its deterministic fallback on every commit. A system that has never run without its provider finds out what breaks during an outage. This one already knows.

The data plane is self-hosted end to end

PostgreSQL 16 and 18 with pgvector. Redis. MinIO for S3-compatible object storage. Docker Compose for single-node, Terraform and Kubernetes for cluster. No managed-service dependency sits in a critical path.

INHERIT ships a SQL compatibility layer and a dedicated CI job proving portability against a real postgres:18 container, so the same code runs on an embedded database in a disconnected enclave and on a clustered instance in a datacenter. Portwize ships twelve Terraform files across VPC, compute, database, cache, streaming, storage, and IAM — the whole plane, declared.

Egress is default-deny

Maven declares an explicit egress allowlist as its second security invariant, enforced by scanner at merge. Portwize carries SSRF guards covering private address ranges, cloud metadata endpoints, DNS rebinding, and non-HTTP schemes. INHERIT scrubs credentials from every inbound message before it reaches a database row, a log line, or the prompt stream.

Observability is Sentry or Prometheus. Both self-hostable. Both configured by the operator, pointed wherever the operator points them. No telemetry returns to us.

Clean architecture is the reason any of this is possible

Ports and adapters at every external boundary — model, embedder, object store, message transport, isolation substrate. Maven represents its production isolation layer as a clean port with a local adapter behind it, which is precisely why the substrate can be swapped for a customer's own without touching the platform. INHERIT degrades from semantic retrieval to keyword overlap when the embedding backend cannot be imported, rather than failing.

Strict typing throughout. No any, no @ts-ignore, enforced by invariant test. Platform code held capability-agnostic by scanner. The discipline that makes a codebase readable is the same discipline that makes it portable.

Government compliance posture

The model is yours. The weights are yours. The database is yours. The only thing we ship is the code, and the code has one door.

Standalone readiness by system

Transferable capability

For solicitations where the domain differs. The argument is method transfer, not equivalence. Each line below is sourced to one of the five engagements above.

Key personnel

Nicholas McGinnis — Principal.

Delivery and product leadership across every engagement above. Prior officer-level role in a medical-legal technology organization. Named co-author on published technical research in specialized healthcare language models. Panelist on data ownership and the information economy with the co-founder of Rackspace.

Aggregate record