
AI UX
The Canvas Becomes a Control Surface for AI
Figma’s agent, Make, and Weave show how selection, visual controls, and design-system context can help people direct AI while keeping decisions inspectable.
Read article →Opens Majestic Labs to discuss your project.
Expertise is not prompted.
It is trained.
This manifesto is about the PersonaSI approach to doobucheniye: extracting specialist judgment from real work, turning it into training material, and verifying that the resulting agent can decide with competence, benevolence, and integrity.
PersonaSI gives specialists a working AI agent on day one — avatar, discovery engine, orchestration, nodes, and security pre-wired. No prompting marathons. No glue code.
Quantum in process.
Binary in result.
The agent explores all strategies in superposition. The output is always binary — deal closed, or not. This is the structure PersonaSI applies to everything: the partnership philosophy in working code.
— PersonaSI · Core philosophy · Tel Aviv, 2026
The CEO layer is the product speaking as a company: specialist agents, fine-tuning, fingerprinting, deployment, and verified outcomes. It is not philosophy alone; it is the commercial surface of the system.
Hero tagline
Specialist's AI agent — fine-tuned,
fingerprinted, ready to close the deal.
Operating promise
No prompting marathons.
No glue code.
Product block
Built around a single goal:
close the deal.
Outcome layer
Sale · Diagnosis · Sign-off · Handoff
Escalation · Refusal · Deployment
CTA principle
Your expertise deserves to scale
beyond your calendar.
Maintenance rule
Drive is authoritative.
Notion is operational.
The site is the rendering surface.
PARAM — 01
39%
Productivity gain —
personalized AI vs generic
tools (2026 benchmark)
PARAM — 02
78%
of AI failures are invisible —
confidently wrong,
never caught in production
PARAM — 03
$52B
AI agent market by 2030 —
CAGR 46.3% —
specialist agents leading
A specialist agent is not a prompt wrapper. It is a system that fingerprints expertise, fine-tunes behavior, verifies bias, and turns domain judgment into a deployable interface.
The architecture is built for professional outcomes: sale, diagnosis, sign-off, handoff, escalation, refusal. Every node and guardrail tunes toward outcome, not chat.
PersonaSI is the fine-tuning layer. Looktopus is the marketplace. Bias audit is the gate that decides whether an agent is trusted enough to be listed.
Architecture components
Avatar · Discovery engine · Orchestration
Node graph · Security layer · Bias audit
Product pillars
PersonaSI — fine-tuning layer
Looktopus — agent marketplace
Bias audit — built in
Target specialists
Consultants · Advisors · Operators
Domain experts · Customer success
Fingerprint
Voice & expertise extraction
We map your existing work — decisions, patterns, domain judgement — into a structured dataset that captures who you are, not just what you know.
Fine-tune
Train on your data
The model trains on your patterns specifically. Not averaged across the internet. Calibrated against your standards, your audience, and your decision logic.
Audit
Bias verification
Every agent is tested against the three trust axes — competence, benevolence, integrity — before deployment. Systematic bias is caught before it reaches production or a marketplace listing.
Deploy
Ship & compound
Your agent is live. Each interaction deepens the partnership. The model becomes more useful with every cycle — not less, as generic tools do.
PersonaSI instances — active
Signature · VerificAI · Looktopus
Object Passport
First MVP focus
Signature — Writing & Content Agent
Fine-tuned on voice, reasoning, audience
Marketplace status
Looktopus — pre-trained agents listed
after bias audit clearance
Location
Tel Aviv · Hebrew University collaboration
Israel · 2026
Finding 01
Trust judgments differ
The tested models evaluated trust more rigidly than human respondents, with less overlap between ability, benevolence, and integrity judgments.
Finding 02
Bias depends on context
Demographic effects differed across models and scenarios. Evaluation should examine the specific model and task.
Finding 03
Broad experiment, bounded evidence
The study used 43,200 simulations across five models and five scenarios: lending, donations, management, guided trips, and childcare.
Study design
Important limitations
Prompt wording, same-session measurements, and model–human sampling differences limit interpretation and generalization.
PersonaSI implication
Evaluate before relying
Evaluate specialist agents on relevant tasks, inspect demographic sensitivity, and document uncertainty before relying on their judgments. This is a proposed evaluation approach, not a demonstrated product result.
Does the agent demonstrate the relevant ability needed for the domain? Mayer, Davis and Schoorman use the original term “ability.”
Proposed design: evaluate ability on representative tasks with independently defined criteria.
Does the agent act in the client's interest, not just pursue outcomes? Does it know when to pause versus close?
Proposed design: evaluate whether outcomes remain aligned with the user's interests.
Does the agent adhere to principles the client finds acceptable? Is the judgment consistent and auditable?
Proposed design: document policy adherence, uncertainty, and demographic sensitivity.
The tested model judgments were less correlated across the three dimensions than human judgments. How a product should aggregate them remains a design choice requiring validation.
Holistic aggregation and integrity gates are proposed choices, not established requirements.

AI UX
Figma’s agent, Make, and Weave show how selection, visual controls, and design-system context can help people direct AI while keeping decisions inspectable.
Read article →
Emerging Interfaces
New AI interfaces make room for judgment through task-specific controls, contextual feedback, well-timed checkpoints, and inspectable demonstrations.
Read article →
Learning Interfaces
Three recent studies offer practical lessons for the interfaces people use to train specialist agents.
Read article →
AI Learning
Personal agents need a disciplined way to decide what a correction should change. Recent research offers approaches to selective training, editable memory, and evaluating whether feedback produced a meaningful improvement.
Read article →
Memory
Personal agents should prove that experience improves later work while avoiding new mistakes. Recent benchmarks point toward evaluating transfer, retention, appropriate use of personal context, and the effort required from users.
Read article →Research library
Continue with prospective memory, evaluation drift, and evidence-based trust.
View all research →A proposed discovery role could collect relevant evidence, an orchestration role could organize it, and a verification role could test whether judgments remain consistent across relevant variants. These roles require implementation and validation.
Competence failure
A plausible response
can still be wrong.
Benevolence failure
An outcome can be pursued
at the user's expense.
Integrity failure
A judgment may differ
across demographic groups.
Scenario set
Run domain trust simulations
Credit decision, recommendation, qualification, escalation, and handoff scenarios are adapted to the specialist's domain.
Variant test
Hold trust signals constant
Demographic attributes vary while competence, benevolence, and integrity signals stay identical. The decision delta becomes the bias signal.
Audit report
Measure demographic variance
Proposed output: document scenario coverage, demographic variance, bias flags, uncertainty, and review status.
Listing gate
Define an admission rule
Proposed design: a marketplace admission rule could require documented evidence and review before listing. This is not a demonstrated current gate.
PersonaSI exists to make specialist knowledge operational: fingerprinted, fine-tuned, bias-audited, and deployable as an agent that can be trusted in real decisions.
Category
Specialist Intelligence
Contact
Principle
Verification over prompting.
Judgment over chat volume.