Cognitive memory · agent infrastructure · AI product engineering

I build AI agents that remember, learn, and act without surrendering human control.

I design persistent agent systems whose memory, evidence, authority, and actions remain inspectable. My work spans governed cognitive memory, multi-agent orchestration, deterministic authorization, and full-stack AI products.

In practical terms: I build AI systems that can accumulate experience over time while preserving where information came from, handling conflicts safely, restricting what agents may do, and keeping consequential decisions reviewable by humans.

Pittsburgh, PA · open to agent infrastructure and AI systems opportunities
56 worldsseparately seeded frozen holdout
12 testsadversarial authorization cases in public
9 operationsgoverned Hermes memory tools
2 repospublic source and runnable evidence
One coherent thesis

Persistent agents need more than a capable model.

They need memory that can change without losing provenance, teams that can coordinate without sharing ambient authority, deterministic controls below model output, and products that expose uncertainty instead of hiding it.

I am exploring agent infrastructure, AI systems, and applied AI engineering roles. Review the public source or ask for a private systems walkthrough.

Featured work

Six systems, ordered by what a reviewer can verify.

Status, ownership, evidence, and limitations are explicit. Public repositories and live surfaces are linked directly; private implementation details are labeled rather than implied.

ImplementedPresent in code or a working system.
MeasuredBacked by a linked protocol or artifact.
ExperimentalA hypothesis with a stated evidence limit.
PrivateAvailable only in a qualified walkthrough.

Cognitive Memory Fabric

Flagship · open source · alpha research system · 2026
PUBLIC SOURCE

Problem

Persistent agents cannot safely treat memory as a bag of embeddings. Beliefs change, sources disagree, old facts become obsolete, and useful abstractions require more than a plausible model completion.

What I built

A local-first cognitive memory substrate with authoritative SQLite state, an append-only event ledger, provenance and evidence, temporal cognition, contradiction preservation, consolidation, reversible forgetting, an Obsidian projection, symbolic replay, and opt-in neural reranking.

Governance decision

Models may propose an abstraction or replay candidate. Deterministic policy decides whether evidence is sufficient. Neural output may rerank the same symbolic candidate set; it cannot add candidates or mutate lifecycle state.

Hermes integration

A first-class provider performs bounded automatic recall. Nine explicit tools make durable changes inspectable, while returned memories remain untrusted evidence with provenance, confidence, validity, conflict, and supersession state.

Governed memory lifecycle
ExperienceProvenance-bearing encoding and temporal organization
EvidenceConfirmation, contradiction, source monitoring, and confidence
ReplaySymbolic retrieval with optional bounded neural reranking
ConsolidationEvidence-gated principles and guarded identity inference
ForgettingArchive, supersede, and restore instead of silent deletion
36,864LIF neurons in the GPU-oriented replay model
770,048sparse excitatory and inhibitory synapses
5 conditionsbasic through fabric plus neural replay
Publicsource, tests, protocol, reports, and artifacts
Measured result and boundary

On the frozen 56-world holdout, symbolic replay reached 96.43% and the frozen neural candidate reached 98.21%, helping one rank and harming none. The result was directionally positive, but the paired interval touched zero and the preregistered 3-point superiority threshold was not met. No superiority claim is made.

This is an alpha research system. The neural subsystem is a hippocampal-inspired engineering model, not an anatomical, biological, or clinical claim.

Hermes Agent Lab

Independent · active private lab · native Windows · 2026
ACTIVE LAB

What it owns

Execution, orchestration, and authority. A manager routes work across isolated role profiles with separate models, tools, state, concurrency, and permissions. Durable task dependencies and context packets preserve continuity.

What CMF owns

The memory lifecycle. First-class Cognitive Memory Fabric integration adds bounded recall, explicit durable writes, evidence-aware revision, temporal context, reversible archival, and an inspectable Obsidian projection.

Concrete workflow

A versioned revision moved from manager to isolated reviewer, returned for changes, produced exactly one corrected revision, passed dependency-bound verification, and then reconciled idempotently without gaining live execution authority.

Failure design

Unknown and malformed states fail closed. Finite retries, idempotency, atomic state, encrypted backups, hash-bound artifacts, health checks, rollback procedures, and skill supply-chain review are designed into the lab.

Execution and memory boundary
ManagerPlans, delegates, and escalates by risk
Isolated rolesIdentity, state, model, tools, and authority stay separate
CMF providerBounded recall plus nine explicit durable-memory tools
Human gateEvidence and review precede consequential action
First-classprovider participates in Hermes lifecycle
9 toolsexplicit creation, evidence, archive, context, and repair
6/6 probessanitized gateway snapshot from July 21, 2026
Private labpublic integration contract and workflow evidence

Public evidence covers the integration contract and a disposable orchestration workflow. It does not certify continuous health, autonomous production operation, or a completed disaster-recovery restore.

Agentic Systems Showcase

Public clean-room reference · Python · security architecture
RUNNABLE EVIDENCE

Problem

A model can propose a useful action, but prompt text is not authorization. Tool availability, tenant scope, approval, and replay protection must remain below model output.

What is inspectable

A dependency-free authority-boundary reference with default-deny tools, tenant and scope enforcement, action-bound short-lived confirmation, argument-change invalidation, replay denial, and fail-closed unknown tools.

Proof

Twelve adversarial authorization tests, CI, architecture decision records, sanitized system cards, and the portfolio source establish an inspectable public evidence boundary.

Known limits

Replay state and confirmation keys are process-local, approver identity is trusted input, there is no durable audit store or provider integration, and distributed concurrency is outside the demonstration.

Live invite-only alpha · end-to-end owner

Wadevo Finance

Cash-flow forecasting, bills, credit tracking, and natural-language search. A deterministic attention engine owns alerts and projections; bounded AI interprets scoped questions. The public landing and authentication boundary are live, while source and product access remain private.

Open live surface ↗

Live constrained sampler · end-to-end owner

Wadevo Design

An AI-assisted brand exploration product with public examples and guided exploration. Authenticated studio, settings, generation APIs, and spend-bearing model routes stay protected behind a fail-closed boundary.

Try the sampler ↗

Access-controlled prototype · end-to-end owner

Healthspan

An iPhone, watchOS, API, worker, and web system that exposes provenance, coverage, baseline, confidence, algorithm version, and limitations for derived wellness metrics. It is not a medical device or clinical claim.

Request walkthrough ↗

Additional builds

Other ways I have applied the same boundaries.

Prototype · Jul 2026

CMA Copilot

Local-first, synthetic comparable-sales workflow with explicit conflict resolution, editable pricing strategies, evidence review, and a human-owned recommendation.

Desktop application · Jan 2026

Resume Builder

Rust and Tauri application for artifact-to-resume generation, job tailoring, gap analysis, interview preparation, and multi-format export.

Web application · Feb-Jun 2026

Wadevo Tasks

AI-assisted task organizer that converts an unstructured brain dump into actionable work for an ADHD-friendly workflow.

Visible implementation demo

Evidence designed to be inspected, not merely asserted.

Claims are structured, project status is explicit, public and private proof are separated, and experimental results carry their limitations beside the numbers.

Machine-readable files mirror the human page instead of introducing a more flattering version. Public read-only endpoints permit cross-origin retrieval for browser-based agents without relaxing the rest of the site's isolation policy.

Two public repositories make the strongest claims inspectable: Cognitive Memory Fabric for governed memory and experimental replay, and Agentic Systems Showcase for deterministic authority boundaries.

Engineering approach

Three rules that connect the stack.

Memory is governed state

Provenance, contradiction, temporal validity, consolidation, archival, and restoration are lifecycle concerns, not retrieval tricks.

Models propose, systems verify

Models interpret, search, rank, and draft. Deterministic code owns evidence gates, permissions, invariants, and mutations.

Humans retain authority

Consequential action follows explicit scope, review, confirmation, recoverable state, and a visible failure path.

Hiring

Building agents that need memory without losing control?

I am interested in agent infrastructure, AI systems, and AI product engineering roles. I share private implementation walkthroughs with qualified hiring teams.