Changelog

Every meaningful change to iMario, in reverse chronological order. For the story behind 1.0, read the launch post.

1.2 May 11, 2026

Persona Designer v3 and a sharper DG Expert

A ground-up rewrite of the two agents you talk to most. Briefing a cohort and shaping a Discussion Guide now feel like working with someone who actually heard you.

Persona Designer, rebuilt

Intent classification on every turn. The agent now reads your brief through a dedicated request classifier and routes between new cohort, refine existing, narrow filter, and clarify paths, instead of guessing.
Standard questionnaires built in. Demographic, psychographic, and category-specific question banks are bundled, so an audience like "15 first-time SaaS founders deciding on analytics tools" arrives pre-instrumented with the right axes.
Filter and individual extractors. Free-text constraints ("based in Germany, never used Mixpanel, switched tools in the last 12 months") are parsed cleanly into structured filters and applied during sampling, not after.
Conversational creation UI. New CreationChoiceCard, LinkedInImportCard, PersonaDesignCard, and PersonalityTestCard turn the persona chat into a guided flow with one click per decision.

Discussion Guide Expert, with constraints

Constraint-aware drafting. The DG agent now respects per-question constraints (length, response type, must-cover topics, forbidden phrasing) end to end, with a new test_dg_constraints.py suite guarding regressions.
Tighter follow-up logic. Follow-up generation now references the live persona profile, so probes land on what this respondent actually said, not the model's defaults.

Prompt-injection defenses

Canary-token framework. A new core/security/canary module plants tracer tokens through prompts and detects exfiltration attempts in agent output, with test_canary_attacks.py covering the common vectors.

1.1 May 8, 2026

Wallet ledger, Deep Dive analysis, and brand-styled exports

Two heavy lifts under the hood: a real wallet for credits, and a Report Engine that finally exports in your brand, not ours.

Wallet ledger for credits

Atomic, idempotent, race-safe. Every credit mutation now goes through wallet_service, with SELECT FOR UPDATE row locks, an immutable credit_transactions ledger, and idempotency keys so Stripe webhook replays never double-charge.
DB-level safety net. A CHECK (balance >= 0) constraint in Postgres rejects overspend at the storage layer, even if application logic has a bug.
Cleaner refunds. LIFO refunds across multiple active subscriptions now reconcile against a single source of truth instead of three legacy columns.

Per-Question Deep Dive (opt-in)

Report node's new Deep Dive toggle turns per-question analysis into a billable, opt-in pass with its own pipeline (per_question_pipeline), instead of riding as a free byproduct of the report.
DeepDivePreviewCards and DeepDiveUpgradeModal give the reader a taste of what Deep Dive surfaces (sentiment splits, dimensional themes, representative quotes) before they enable it.

Brand-styled report exports

Design Tokens engine. Upload a brand reference (.pdf, .docx, or .pptx) and extractors pull your fonts, palette, spacing, and section styles into a token set the renderer uses for every export.
Export cache. Re-exporting the same report at the same theme no longer regenerates from scratch, courtesy of the new export_cache table and migration.
Audit-grade chapters. A new chapter renderer and test_audit_grade_chapters.py ensure findings, evidence, and limitations stay aligned across markdown, PDF, and DOCX.

The biggest release in iMario's short history, and the one that finally makes the product match the vision. Three months of MVP testing, rebuilt from the studs. Read the full story in Introducing iMario 1.0.

A brand new cognitive model for Synthetic Individuals

Deep Identity. Every individual ships with a full narrative profile: background, values, behavior patterns, communication style. Identity is a structured model that constrains every response, not a paragraph in a system prompt.
Expert Reflection. Every response is silently cross-examined by a panel of built-in experts (demographer, psychologist, economist) who surface hidden motivations, price anchoring, status signaling, and what the respondent left unsaid.
Knowledge Boundaries. Dr. Chen the cardiologist answers heart-failure questions with clinical precision and politely declines to opine on your crypto thesis. No more hallucinated expertise.
Human-like memory. Working, episodic, and semantic layers, each decaying along an Ebbinghaus forgetting curve. Sarah remembers last week's conversation, vaguely recalls one from two months ago, and has forgotten this morning's small talk entirely.

Three ways to create an audience, in minutes

One sentence, any audience. Type a brief in plain English ("30 mid-market CFOs, half happy with their current tools, half actively looking to switch") and iMario generates a diverse, statistically grounded cohort. Distribution-first sampling, diversity validation, and evolutionary optimization run under the hood, so you never get 30 clones wearing different hats.
One LinkedIn URL, one exact person. Paste a profile and iMario builds a synthetic individual anchored in that person's real career arc, domain knowledge, and communication patterns.
One assessment, a synthetic you. Take an 8-minute personality assessment and iMario creates a synthetic twin of yourself, ready to represent you in meetings and respond to requests the way you actually would.

A canvas, not a funnel

Tasks are first-class objects on a visual node editor. Drag any content (a Discussion Guide, a landing page, an email draft, a pricing table, an API request), connect it to any audience, connect to the output shape you want, and hit run.
A general-purpose execution surface, not a research tool with a fancier UI. The same canvas covers Discover, Validate, Launch, Optimize, Scale, and Build use cases.
Parallel by default. Every task fans out across your audience in parallel, dropping turnaround from weeks to minutes.

Agent Skills (preview)

Equip Synthetic Individuals with executable skills (send an email, review a document, join a meeting, hand off to a teammate) and let them work for you inside your existing tools. Pipe them into production AI agents over the API to ship products with a real human layer on top.

Want to follow what we ship next? Reserve a spot on the waitlist, or send feature requests via Feedback.