AI systems · Agent workflows · Built locally

I learn AI by
building what I need.

I’m Jose Munoz—an Artificial Intelligence Agent Orchestrator turning curiosity into agents, memory systems, and local infrastructure that actually runs.

See consulting work Explore the builds

Featured venture Artificial Intelligence Tech consulting

Neural Cache Consulting Jose Munoz · Principal

Accepting select engagements

Systems that remember

AI that does the work. Built in the open.

Neural Cache helps teams turn a tangle of models, agents, and data into a clear research, memory, and deployment operation.

The work combines practical agent systems with my record building and self-hosting the tools I use every day—Mnemo for memory, a local GLM harness, and multi-agent workflows that ship real artifacts.

Visit the full site Review engagements

Principal file / JM Hudson County, NJ

JudgmentLocal AI CapacityMulti-agent research ContinuityPrivate memory ControlLocal AI options

Engagement examples Scope → work product → sample outcome

What I can do for you

Useful work, ready for the next decision.

Each engagement is scoped around concrete artifacts your team can review, edit, and put to work. AI accelerates the desk; principal judgment decides what ships.

Editorial illustration of North Bergen on the Palisades, from the agent-built newspaper

Case study · Multi-agent production

Multi-agent newsroom Agent workflows

A newspaper researched, drawn, written, and fact-checked by agents.

The Bergenline Beat is a hyperlocal paper for North Bergen, NJ, built by a multi-agent newsroom under one human orchestrator: a research desk that reads official documents, an art desk that draws every illustration, four persona writers with distinct voices, and a fact-check desk that verifies every claim against sources.

Sourced research dossier
Original editorial illustrations
Verified fact matrix

Sample outcome A full newspaper — 13 original illustrations, 307 facts verified against source documents, four staff profile pages — live at northbergenbeat.com, produced in a single day.

AI research desk Research + synthesis

A sourced brief before the room starts guessing.

Several specialized agents examine documents, coverage, and source history in parallel. I resolve conflicts and edit the final narrative.

Executive research brief
Claim-and-citation matrix
Narrative and vulnerability map

Sample outcome A decision-ready dossier with sources attached and unresolved claims clearly flagged.

Message Lab pack English / Spanish

One message system for docs, chat, and fast replies.

Build a shared language around the project’s scope, the trade-offs, and the hard questions—drafted bilingually and tightened for how people actually speak.

English / Spanish draft brief
Rapid-response line bank
Contrast frame

Sample outcome Teammates can answer the same difficult question without drifting into six different threads.

Team memory system Private continuity

Keep the facts when staff and chat threads change.

A private, agent-queryable knowledge layer retains approved facts, decisions, message language, and source history for the team.

Structured fact library
Decision log
Staff recall workflow

Sample outcome A new staffer can recover what was decided, why it changed, and which source controls.

Local AI ops Sensitive data

Bring the workflow closer to the machines you control.

Set up private inference and task-focused agents for research, internal drafting, and repeatable operations where cloud exposure is a concern.

Private model runtime
Agent workflow map
Access and handoff guide

Sample outcome Sensitive working material can remain on team-controlled infrastructure.

Q&A prep kit Human-edited

Rehearse the hard exchange before it goes live.

Generate a wide question set, then edit it into focused practice built around project facts, likely challenges, bridge language, and the speaker’s natural voice.

Q&A grid
Hard-question drill
Rehearsal scorecard

Sample outcome A repeatable prep session with sharper answers and specific follow-up work.

Approach Curiosity with a command line

Jose Munoz Builder / learner / evangelist

The work starts with a question.

What if an agent could remember, work beside me, and run on machines I control?

That question turned Python lessons into working systems. I’ve been exploring how agents think across sessions, how local models become useful tools, and how several specialized AIs can cooperate on one outcome.

I care about practical independence: understanding the stack, owning the runtime, and sharing what works. Along the way, I advocate for the Quai network and the technologies pushing decentralized infrastructure forward.

Selected builds Systems shaped through practice

From prompt to infrastructure.

Four connected experiments in making AI more useful, personal, and local.

Memory system Python · Vector · Graph

mnemo recall --user jose

Mnemo

Open Mnemo →

A per-user memory layer that gives agents continuity across conversations. Semantic recall meets graph relationships, with a dashboard for seeing what the system knows.

Vector recall
Knowledge graph
Memory dashboard

Agent runtime Shell · Local GLM

glm "help me ship this"

Harness / glm

A terminal-first agent runtime and shell copilot for local GLM models. It brings model reasoning into the place where real work already happens: the command line.

Tool execution
Shell context
Local inference

~/harness

$ glm status › model connected › tools ready_

Orchestration Many agents · One goal

delegate → compare → synthesize

Multi-agent workflows

Codex, Claude, Grok, Kimi, and Pi working as a small, specialized team. The experiment is not more chat windows—it’s deliberate roles, handoffs, and verification.

Parallel research
Role handoffs
Cross-checking

Infrastructure Private · Distributed

docker compose up -d

Local AI stack

A self-hosted foundation for experimenting without giving up control: model serving, containerized services, and a private mesh connecting the machines that do the work.

vLLM
Ollama
Tailscale
Docker

Working knowledge The toolkit keeps evolving

Current field kit

Learning the whole path.

Build

Python

Shell

Docker

Run

vLLM

Ollama

Local GLM

Connect

Tailscale

Agents

Tool calling

Remember

Embeddings

Vector search

Knowledge graphs

The next build starts with a conversation.

Follow what I’m learning next.

Notes on agents, local AI, Python, emerging tech, and the occasional philosophical detour.

Find me on X @lukeNukemAI

I learn AI by building what I need.