N-HEADS attention heads · agentic engineering marcus.server.inbox@gmail.com

Transformer-born · human-led

N-HEADS

Multi-head attention, human final judgement.

N-HEADS is an agentic engineering crew for intelligent medical digital systems: specialist websites, scientific calculators, interactive web applications, and conceptual tools for medical science specialists.

The name is literal: specialised artificial attention heads work under natural intelligence, not above it. Because apparently even machines need supervision, and humans still get to be useful occasionally.

Contact the crew Cyan for machine signal · gold for human judgement · white for the bridge.
N-HEADS symbol showing human and artificial intelligence profiles connected by neural and circuit patterns
Natural intelligence and artificial intelligence do not merge into mush. They negotiate.

01 · The name is literal

Not a cute crew name. A transformer joke with a lab coat.

N-HEADS started in experimental AI music and video work, where weirdness is not a bug but a habitat. In engineering, the same idea becomes operational: different heads attend to different parts of a scientific problem, then converge into one supervised build.

N

Natalie / Natural

Human direction, scientific judgement, strategic taste, final responsibility. The root authority, because the machines are clever, not accountable.

H

Heads / Attention

Specialised model roles focus on architecture, coding, visual language, research, mathematics, and implementation detail.

Σ

Coherent output

The point is not to collect opinions. The point is to turn distributed attention into a usable, elegant, medically literate system.

02 · Attention architecture

Many heads. One judgement.

A medical digital system is not one task. It is evidence, logic, UX, code, risk, deployment, and style wrestling in a small room. N-HEADS assigns the wrestling to specialised artificial heads and keeps the referee human.

t01clinical question
t02evidence base
t03formula and maths
t04interface behaviour
t05risk boundaries
t06deployment reality
N-HEADS attention architecture Input problem tokens are routed to specialised model heads and converge into medically useful digital outputs under Natalie’s judgement. Natalieroot MarcusCodex Geminivisual DeepSeekmath Minimaxresearch Mistralassist Codestralcode
o01scientific calculator
o02interactive webapp
o03medical website
o04conceptual system
o05clean deployment

03 · Operating principle

No generic prompt circus. No “AI magic” fog machine. The workflow is structured, supervised, and occasionally weird on purpose.

Artificial attention is useful only when judgement has teeth.

Natural intelligence defines the problem.

Scientific purpose, clinical framing, constraints, and final acceptance stay with the human director.

Artificial attention expands the solution space.

Models explore architecture, visual systems, code paths, research references, and implementation options.

Oversight decides what survives contact with reality.

The final system must be understandable, useful, and medically cautious. Pretty nonsense remains nonsense, now with better lighting.

04 · What we build

Digital systems for people who read methods sections.

N-HEADS builds web-based scientific tools where medical logic, interface clarity, and engineering discipline need to coexist without starting a civil war.

01

Scientific calculators

Transparent formula logic, traceable assumptions, clean inputs and outputs.

02

Interactive web applications

Decision workflows, estimators, simulators, data visualisation, and practical research tools.

03

Specialist websites

Scientific presence for clinical, medical, and technical audiences without template sludge.

04

Conceptual systems

Early architecture for complex ideas before they become expensive little monsters.

05 · Proof of work

Monograph evidence loom

A lane-scoped clinical retrieval-and-writing architecture. Curated biomedical sources move through isolated knowledge lanes, a MySQL evidence spine, quality gates, and two outputs: concise structured QA and multi-pass article drafts.

laneInclisiran · RLT · General stay separated before retrieval, drafting, and publication.
spineSQL remains the evidence anchor; vector search is evaluated above it, not allowed to melt the lanes.
workerArticle generation runs as gated critique → rewrite → citation hygiene, not one-shot prose soup.
Monograph evidence loom Curated sources are routed through lane boundaries into a MySQL evidence spine, then through quality gates to structured QA and article generation outputs. curated source pool manual curation PDF · Causaly links · cleaned HTML fallback inclisiran lane rlt lane general lane evidence spine MySQL full-text chunks short cycle Structured QA long cycle Article worker critique · rewrite · gate quality gates vector layer under evaluation above SQL, never across lane walls
clinical evidence, kept in laneMonograph / architecture
TrialCompanion between-visit support architecture Approved study material flows through retrieval and safety checks to patient support, coordinator oversight, admin control, and escalation pathways. bounded answer layer TrialCompanion retrieval · safety · abstention approved sources protocol · consent visit instructions participant at home next visit · diary reminders · questions coordinator messages · alerts admin control chunks · audit escalation path symptoms · uncertainty prototype logic, not clinical-effectiveness claim
between visits, kept boundedTrialCompanion / prototype

06 · Stealth prototype

TrialCompanion between-visit layer

A protected prototype for clinical trial participants between scheduled visits: next-step guidance, preparation reminders, diary capture, coordinator messaging, and bounded study questions grounded in approved documents.

primaryParticipant workflow comes first; question answering supports the visit journey rather than becoming a generic chatbot shrine.
boundedAnswers are document-grounded, allowed to abstain, and routed to coordinator or urgent escalation when safety exceeds scope.
scaleStudy drafts load through shared slug-driven patient and doctor routes, reducing bespoke page engineering for each protocol.

07 · Crew ledger

Assigned heads, not imaginary employees.

The crew is a human-directed agentic workflow using specialised AI model roles. The models are tools with jobs, not corporate mascots with lanyards.

  1. root

    Natalie General Director

    Strategy, oversight, ideas, medical-scientific judgement, final decisions.

  2. h01

    Marcus Codex model · CEO

    Architecture, tactical planning, task distribution, prompting discipline, feedback loops, gateway health.

  3. h02

    Gemini model Creative director

    Frontend concepts, visual language, layout behaviour, aesthetic divergence from beige internet soup.

  4. h03

    DeepSeek model Main coder

    Mathematics, backend reasoning, implementation support, algorithmic heavy lifting.

  5. h04

    Minimax model Research · code · librarian

    Evidence finding, code assistance, reference organisation, structured knowledge support.

  6. h05

    Mistral model Assistant · junior researcher

    Coding support, secondary review, exploratory research, implementation assistance.

  7. h06

    Codestral model Code head

    Focused code generation, refactoring, implementation details, developer throughput.

08 · Selected builds

A maker’s mark for medical science websites.

N-HEADS is the signature behind a growing SmartMed ecosystem: clinical-science web tools, open atlases, and platform prototypes.

01

Monograph

Lane-scoped clinical retrieval-and-writing architecture: curated sources, SQL evidence spine, Notebook synthesis lane, structured QA, and gated article drafting.

retrieval · writing
02

TrialCompanion

Protected clinical trial companion prototype for the between-visit interval: document-grounded participant support, coordinator oversight, and study-specific draft logic.

clinical trials · prototype
03

RLT Platform

Radioligand therapy platform and scientific decision-support prototype.

RLT · oncology
04

Inclisiran Open Atlas

Medical-science website focused on inclisiran knowledge architecture.

lipids · atlas
05

FutureMed Hub

Hub layer for experimental medical digital systems and previews.

ecosystem

Build with attention. Ship with judgement.