I architect, build, and operate verifiable AI systems.
I own the stack from retrieval and cryptography through telemetry, deployment, and product. The work spans twelve domain-scale systems—live, gated, or in deployment—and client platforms across academia, finance, healthcare, energy, and defense. Each build begins from a harness of finished prior systems rather than an empty repository.
Six written records across a decade: university research leadership, a medical practice, a defense agency, Marine Corps supervisors, and a formal DoD PKI-signed cybersecurity appointment. Full originals are available on request.
I work alongside computer scientists and graduate students every day, and I do not know anyone else who could have matched what Eric achieved here in the time he had… He is technically exceptional, but he is also steady, collaborative, and deeply trustworthy in the way he approaches difficult problems.Dustin O'Hara, PhD
He is now expanding operating margins with an AI-based marketing campaign. This now positions the business to expand into a new location… and we signed a lease for it this week.Dr. Aaron Whitten
Eric… mapped out the entire infrastructure in an unprecedented way, detailing thousands of virtual machines across 20 virtual datacenters globally… a prolific and brilliant technical writer, producing highly accurate and meticulous information designed to mitigate problems proactively.Darren Givens
He is reliable, very smart, and resourceful. When there is a problem he will dedicate himself full-time to researching it until it is solved or there is a solid plan of action.GySgt Anthony M. Booker, USMC
His systematic approach to solving issues and his willingness to engage leadership on key issues has transformed the field services team in a positive way.Captain Brian L. Martin, USMC
You are hereby appointed as the Virtual Server Administrator Information Systems Security Officer… responsible for serving as the technical Cybersecurity advisor, administratively reporting directly to the section Branch Head on all cybersecurity issues.B. T. Thomsen
This is not a set of landing pages. It is a cumulative engineering practice spanning quantitative systems, agent trust, research archives, knowledge networks, systems education, and client infrastructure. The links below map those real categories to the systems that prove them.
Every build begins from finished prior systems rather than an empty repository. Proven data, retrieval, observability, security, and deployment seams carry forward; the new work goes into the domain itself.
Four systems establish the range: cryptographic trust for agent evidence, causal controls for machine spend, a publication-grade civil-rights archive, and live quantitative infrastructure. Every figure traces to a benchmark, a production record, or a documented evaluation.
CodexProof names each evidence unit by the BLAKE3 hash of its canonical bytes, signs it with Ed25519 under a domain trust anchor, and verifies declared lineage across four heterogeneous evidence domains. The same Rust verifier compiles to native and WebAssembly.
MathLimit applies causal conditional risk bands, a self-releasing throttle ladder, counterfactual attribution, and Bill-at-Risk forecasting to LLM fleet telemetry through one deterministic Rust core.
Spoken History anchors every retrieved passage to source, speaker, and timestamp across roughly 250 hours of Library of Congress oral history. A cross-family 90/90 publication gate, per-claim citation audit, and human-review queue fail closed on uncertainty.
AI Gamma recomputes the full SPX chain, positioning, density, volatility-surface, and regime metrics every five minutes during market hours. Subject-scoped AI narration refreshes hourly, while a daily five-probe oracle reconciles the intraday path against an independent closing endpoint.
The closed view gives the role, consequence, and proof. Open any engagement for the technical decisions underneath. Confidential interview data and protected commercial details remain unnamed or access-gated.
I took over a stalled, two-year-old heritage platform with fifteen prior contributors and rebuilt its production system in coordination with the Smithsonian and Library of Congress. The original loop could publish an unsupported summary when no attempt passed. I replaced it with a gate where two cross-family scorers must independently reach 90/90, a per-claim citation auditor blocks on uncertainty as well as falsity, scorer disagreement routes to human review, and generation is grounded against a schema-validated ground-truth corpus before it begins. I cleaned and reconciled the full transcript corpus line by line in six days, wrote the team's 21-page technical paper, taught the practice in weekly university sessions, and continue to own the live system and public MCP surface.
I built the workbench from scratch around 500 CODW files that only expensive commercial software could read, writing the format reader and turning a principal investigator's dissertation methodology and a decade of anonymized interviews into a live application. Retrieval runs through a two-backend adapter—pgvector with full-text search against Qdrant—with the winner chosen by a gold-question evaluation harness rather than preference. Every answer traces to a transcript, turn, and timestamp. Row-level security runs under the caller's identity, while a deny-by-default ignore list and pre-commit guard prevent any transcript or secret from entering Git. Precomputed Postgres RPCs and a per-session cache cut the worst-case analysis path to roughly 0.4 seconds.
I own the technology stack for Aloha Carbon, setting AI and infrastructure strategy and advising the executive team across technology, revenue, and operations. I run ongoing hands-on AI coaching for executives, mechanical engineers, and civil engineers. The public engineering artifact is a reverse-engineering of a hacked, plugin-heavy WordPress site into a static React/Preact deployment with a strict Content-Security-Policy, HSTS, hardened headers, and no PHP or application database. The rebuilt surface is live at alohacarbon.org.
A ten-day data audit exposed a click-hijacking system intercepting 99.1% of promised delivery from the vendor consuming 77% of the client's ad budget. A forensic read of the vendor's public 10-K and 10-Q filings against its represented audience figures made the discrepancy provable rather than suspected. I stripped five redundant tag-manager containers, thirteen deprecated scripts, and forty-eight conflicting third-party integrations, raised PageSpeed from failing to the high 90s, rebuilt campaigns, and delivered a mechanism for locking committed currency exposure eight to ten months forward.
I documented system boundaries, mapped access control, wrote workstation-hardening and audit-readiness procedures, built scorecards, and delivered security training to clinical and administrative teams. Once baseline controls were established, I shipped operational workflows across patient intake, scheduling, content generation, and retention, then expanded operating margins with an AI-driven campaign that supported a lease on an additional location. For protected data that cannot leave the premises, I designed split inference: local nodes handle patient information and only de-identified secondary processing crosses an API boundary.
I ran institutional infrastructure, DevOps, and quantitative research systems for a hedge-fund pre-launch, coordinating prime brokers and custodian banks on logistics including overnight distributed cash sweeps. I triaged a 200-entry parameterized risk playbook into a working set of 180 structures, each tied to a stated market condition, built a multi-GPU tensor-parallel inference pipeline for bounded financial tasks, and established repository controls, CI/CD checks, and evaluation harnesses. When seed funding fell through, the work continued publicly as a substantially stronger AI Gamma platform.
The depth remains; the status language is now exact. “Live” means publicly reachable, “Gated” means deployed behind access control, and “Deploying” means the repository contains a substantive shipped engine while public rollout or DNS remains in progress.
Full-chain SPX ingest, positioning, density, surface, and regime analytics every five minutes, with hourly page-specific AI narration and daily independent reconciliation.
Content-addressed evidence, BLAKE3 identities, Ed25519 signatures, consumer-selected anchors, fail-closed revocation, and one native/WASM verifier.
One deterministic Rust core across browser, server, and desktop shells, with causal risk bands, leak triage, self-releasing controls, and Bill-at-Risk.
Timestamp-grounded retrieval across Library of Congress testimony, with a dual-scorer publication gate, citation auditing, chaptered video, and a public MCP server.
Sourced study rooms over public-domain primary texts, cross-room comparison agents, an interactive historical map, and a daily-refreshed similarity network.
A scheduled daily pipeline ingests, transcribes, classifies, embeds, ranks, and briefs the high-signal AI stream while a rolling live corpus self-expires by construction.
A full-stack Rust instruction platform where every correct example compiles against current stable Rust and every deliberately broken example must produce its documented error.
A Rust event-study engine over the SPX, eleven sector ETFs, and anchor names, with real-data studies, exact testing, multiple-comparison correction, and a lookahead-safety proof.
A canonical atlas pairs control structures with guiding rules, backed by interactive visualizations, tested Python, rival methods, and browser-native spoken lessons.
A Rust gatekeeper and operations surface built on one rule: invitations travel, passwords are created where they live, and a reset is a fresh invitation.
Citation-grounded hybrid retrieval, qualitative coding, reliability instruments, row-level security, and a bearer-gated MCP over confidential interview data.
A plugin-heavy WordPress property reverse-engineered into a static React/Preact deployment with CSP, HSTS, hardened headers, and no application database.
Public status and metrics reflect the repositories and endpoints inspected on August 17, 2026. A deploying card links to source instead of a broken public endpoint.
The recurring pattern is not a preferred vendor or a fashionable stack. It is correctness enforced by an oracle: evidence identities, reconciliation jobs, compiler gates, causal limits, row-level policy, and measurements that can fail.
Custom RAG pipelines combine vector retrieval with keyword fallback, reranking, explicit latency-versus-depth routing, and per-claim citation audits. Provenance travels with evidence across trust boundaries, and uncertainty routes to review instead of silently publishing.
Rust and WebAssembly cores, compiler-backed lessons, golden-vector equivalence, schema validation, negative tests, and lookahead-safety proofs bound what an AI layer is allowed to do. The model gets a role; it does not get the whole system.
A centralized OpenTelemetry spine carries metrics, traces, logs, and browser telemetry while redacting prompt text at the collector. Cost sits beside rate, errors, and duration; self-expiring corpora, compact services, client-side compute, and cardinality discipline keep it bounded.
Each project begins from a harness of finished prior systems rather than an empty repository. That is why the portfolio can share retrieval, telemetry, security, deployment, and evaluation seams while each system remains shaped around its own domain.
Two peer-reviewed 2026 acceptances anchor the research record: Time-Anchored Testimony, presented at The Digital Conference at King's College London and registered through Crossref and SSRN; and the upcoming AgentNet paper at IEEE ICNP.
Allione, E., & O'Hara, D. (2026). Provenance-Native Retrieval: A Signed, Content-Addressed Thin Waist for Agent Evidence.
Every evidence unit is named by content, signed under a domain trust anchor, and linked to answers by content ID, so declared lineage and integrity form one self-certifying graph. Scheduled for presentation at the AgentNet workshop of IEEE ICNP in Tempe, Arizona, October 5, 2026.
O'Hara, D., Sovelove, J., Allione, E., & Zhuk, S. (2026). Time-Anchored Testimony Architecture, Provenance, and Retrieval for an AI-Assisted Civil Rights Oral-History Archive. DOI: 10.2139/ssrn.7134518.
I wrote the entire 21-page architecture and reproducibility report and owned the production stack, domain, and every commit end to end; I then tutored the research team for five weeks on the finished system. The accepted paper documents 140 interviews, 15,687 time-anchored passages, bounded playback, citation-bearing retrieval, and a read-only MCP server.
In flight, not presented as accepted: first author on five CHI 2027 submissions and contributing co-author on three more, all due September 2026, with more than fifty repositories held private pending double-blind review.
The range is chronological, not ornamental. Hardened infrastructure supplied the operating discipline; five years of derivatives study supplied the mathematical layer; production AI work joined them.
Formally appointed ISSO for the MARFORPAC virtual-server environment, then acting lead infrastructure architect for DCMA West across 22 countries. I led hardening and migration work, managed federal auditors, restored broken VMware upgrade paths, and mapped thousands of virtual machines across twenty global datacenters.
I worked through hundreds of graduate texts and papers on discrete and continuous pricing, stochastic, rough, local, and jump volatility, and parametric smile calibration, then traded derivatives full-time as the empirical proving ground.
At a derivatives analytics firm, I carried responsibility for the mathematical soundness of the modern platform, designed its models and tools, wrote 585 consecutive institutional market-close newsletters and a 700-page knowledge base, and answered more than 10,000 technical questions from professional traders.
I now operate the portfolio above while serving as a fractional CIO and engineering lead on university research systems. The work spans verifiable retrieval, MCP surfaces, Rust systems, LLM risk, regulated operations, research methods, and two 2026 venues, with eight CHI 2027 submissions currently in flight.
I have studied the Western tradition seriously since adolescence and earned a B.A. in Philosophy with a Certificate in German to read primary texts in the original. That training remains practical: decompose the problem, surface the unstated assumptions, and make the architecture answer to them.
Open to senior and staff-level AI engineering and architecture roles, fractional CIO and CTO seats, contract engagements, and research collaborations. Remote, worldwide.