004CHINVORA AI · INTELLIGENCE DIVISION
Chinvora AI exists to find the product shapes that only artificial intelligence can take — and to build them to the same editorial standard as everything else the studio ships.
// SUB-BRAND · D/03
CHINVORA AI
Chinvora AI is not a separate company. It is the research and product division inside Lucky's Digitals, the way Labs sits inside a larger house: its own identity, its own standards, its own release cadence, but the same operator and the same principles. Everything the division learns flows directly into client engagements, and everything client work exposes — brittle prompts, unmeasurable quality, cost blowouts — becomes a research question here.
The division's working thesis is that most AI products fail for architectural reasons rather than model reasons. A capable model wrapped in an unstructured prompt, with no retrieval discipline, no evaluation harness and no observability, produces a demo that impresses in a meeting and collapses in production. Chinvora treats intelligence as a material with grain and tolerances: something to be engineered, measured and constrained, not sprinkled on top of an existing interface.
Practically, that means every Chinvora project ships with three artefacts before it ships with a UI — a written specification of what good output looks like, an evaluation set that can prove it, and a cost model per interaction. Products that cannot clear those three gates stay in the lab.
001SYSTEM ARCHITECTURE
Most AI products fail for architectural reasons, not model reasons. This is the stack the division holds every engagement to.
Model-agnostic routing across frontier and open-weight models, with per-task selection based on capability, latency budget and cost ceiling. No product is welded to a single vendor, because pricing and capability both move quarterly.
Structured multi-step reasoning: decomposition, critique passes, adversarial review and synthesis. Where a single completion is unreliable, a small ensemble of narrowly-scoped roles is measurably better than a longer prompt.
Hybrid vector and relational memory with explicit provenance. Every retrieved fact carries its source so output can be audited, and stale context is expired on a schedule rather than accumulating quietly.
Written quality specifications, golden datasets, regression runs on every prompt change, and rubric-based scoring. The division treats a prompt edit exactly like a code change: it needs a test before it merges.
The interface layer where most AI products lose their users. Streaming-first UIs, visible reasoning state, editable output, graceful failure and keyboard-native interaction designed with the studio's editorial discipline.
Full-trace logging of prompts, tools, tokens, latency and cost per interaction, with alerting on drift. You cannot improve a system you cannot see, and you cannot price one you have not measured.
002OPEN RESEARCH TRACKS
When does an ensemble of narrow agents beat one long prompt?
Measuring where role-separated agents with adversarial critique actually improve output quality against the cost and latency they add. Early finding: the gain is real on open-ended judgement tasks and negligible on extraction tasks, which is exactly where most products deploy it.
Can every generated sentence carry a verifiable source?
Building citation graphs that survive summarisation, so long-form output remains auditable end to end. The target is a published piece where any claim can be traced to a document, a page and a paragraph without manual reconciliation.
How do you score creative output without a right answer?
Rubric-based scoring with calibrated human sampling, applied to design briefs, copy and naming. The output is not a number for its own sake but a regression guard: proof that today's prompt is not worse than last month's.
What is the cheapest architecture that still clears the quality bar?
Systematic downgrading: replacing frontier calls with smaller models, cached retrieval or deterministic code wherever measurement shows no quality loss. Most AI products can cut per-interaction cost substantially without users noticing anything.
003THE PRODUCTS
A creative ideation engine that argues with you.
MindImagined takes a half-formed idea and pushes it to a defensible brief. Instead of returning a tidy list of suggestions, it runs structured divergence — reframing the problem, generating deliberately distant options — and then subjects each surviving direction to a critique pass that names the assumption it depends on. The output is not inspiration; it is a document you can take into a stakeholder meeting and defend line by line.
// CAPABILITIES
Structured divergence
Forced-distance generation that penalises near-duplicate ideas, so the second and third routes are not restatements of the first.
Assumption surfacing
Every direction is annotated with the belief it rests on and the cheapest way to test that belief before spend.
Multi-agent critique
A critic role attacks each concept commercially and a historian role checks it against prior art, before synthesis.
Brief export
One-click export to a structured brief: problem, audience, constraints, directions, risks, next test.
Session memory
Long-running projects retain their constraint set, so week-four sessions do not re-litigate week-one decisions.
// BUILT FOR
// APPLICATIONS
// ROADMAP
Divergence engine, critique loop and brief export in public beta.
Team workspaces with shared constraint sets and comment threads.
Reference-image grounding and a visual moodboard surface.
Public API so the engine can sit inside other studios' workflows.
// CONNECTION TO STUDIO WORK
MindImagined is the direct product of studio practice: it encodes the way Lucky's Digitals interrogates a brand brief. Client work funds it, stress-tests it and receives it back as leverage — brand engagements that used to need two concepting weeks now need one.
Four agents. One question. A real argument.
The Quadruplet Enigma is the division's core reasoning research made usable. Four permanently-scoped agents — Architect, Critic, Historian and Operator — receive the same question and debate it in structured rounds before a synthesis pass produces a single answer with the dissent preserved. Where conventional assistants collapse ambiguity into confident prose, the Enigma keeps the disagreement visible, because on judgement-heavy questions the disagreement is the useful part.
// CAPABILITIES
Four fixed roles
Architect proposes structure, Critic attacks it, Historian supplies precedent, Operator asks what it costs to run.
Visible dissent
The synthesis names which agent disagreed and why, rather than averaging the room into a bland consensus.
Round budgeting
Debate depth is a dial with a hard token and latency ceiling, so a question cannot quietly cost fifty dollars.
Vector memory per thread
Each debate thread retains its own evidence set, letting a decision be revisited months later with its context intact.
Tool access
Agents can retrieve documents, run calculations and query supplied datasets rather than reasoning from vibes.
// BUILT FOR
// APPLICATIONS
// ROADMAP
Four-agent mesh, round budgeting and vector memory in private alpha.
Custom role definitions so teams can add a domain specialist.
Document-grounded debate over an uploaded corpus with full citations.
Wider beta once cost-per-decision reliably clears the internal ceiling.
// CONNECTION TO STUDIO WORK
The Enigma is where the division's multi-agent research track becomes a shippable surface. Its routing, budgeting and evaluation layers are the same ones Lucky's Digitals deploys inside client AI builds — which is why AI engagements here arrive with evaluation harnesses instead of promises.
The operations agent that runs the studio's back office.
Studio Co-Pilot automates the administrative layer of running an operator-led studio: reading an inbound enquiry, drafting a scoped brief, producing a quote against historical effort data, generating the proposal, and preparing project hand-off documents. It exists because the bottleneck in a one-operator studio is rarely the creative work — it is the six hours a week of scoping, quoting and writing that surrounds it.
// CAPABILITIES
Enquiry triage
Inbound briefs are classified by discipline, budget band and urgency, then routed with a draft reply attached.
Historical estimation
Quotes are generated against logged effort from comparable past engagements, not intuition.
Proposal generation
Scope, exclusions, timeline and payment schedule assembled into the studio's standard document format.
Hand-off packaging
End-of-project documentation, credential checklists and support-window terms produced automatically.
// BUILT FOR
// APPLICATIONS
// ROADMAP
Triage, estimation and proposal generation live internally.
Calendar and invoicing integration; automated follow-up sequences.
Packaged as a product for independent studios, priced per seat.
// CONNECTION TO STUDIO WORK
Co-Pilot is the clearest example of the division's feedback loop: it was built to solve a Lucky's Digitals operations problem, and the effort data it captures makes every client quote across the studio more honest.
A publishing pipeline with fact-checking built into the middle.
Editorial Engine encodes the studio's writing process as a pipeline: research, outline, draft, fact-check, typeset. Its research question is provenance — whether every claim in a finished long-form piece can be traced back to a source document automatically, so publishing at volume does not mean publishing unverifiable text. This is the division's answer to the flood of confidently wrong AI content.
// CAPABILITIES
Citation graph
Every claim is bound to a source span and survives outline and summarisation passes without losing its link.
Fact-check stage
A dedicated pass flags unsupported assertions and blocks publication until each is sourced or removed.
House-voice constraint
Rubric-scored against the studio's editorial standard, so output does not drift into generic marketing register.
Typeset export
Structured output that maps directly onto the site's editorial components — pull quotes, sections, metadata.
// BUILT FOR
// APPLICATIONS
// ROADMAP
Citation graph and fact-check pass under active research.
Full pipeline running the Journal end to end with published provenance.
Offered as a managed service inside studio content retainers.
// CONNECTION TO STUDIO WORK
Editorial Engine is the reason the Journal exists as a real publication rather than a blog. It also feeds the studio's SEO and content retainers, where verifiable claims are the entire value proposition.
004PROMPT LIBRARY
Working prompts for design briefs, architecture reviews, research and marketing — each one version-tracked and rubric-scored before it enters the library.
18 prompts
Intake interrogation, positioning tests, naming sprints and art-direction references.
OPENS TO CLIENTS · Q4 2026
22 prompts
Schema review, failure-mode analysis, refactor planning and handover documentation.
OPENS TO CLIENTS · Q4 2026
14 prompts
Competitive teardown, claim verification, long-form outlining and launch copy.
OPENS TO CLIENTS · Q4 2026
005CHINVORA FAQ
No. It is the intelligence division inside Lucky's Digitals — its own name, identity and release cadence, but the same operator, standards and commercial entity. Client contracts are with Lucky's Digitals; the division supplies the capability.
Yes. AI engagements run through the studio's standard fixed-scope model and start at $3,000, covering agent systems, retrieval architecture, workflow automation and AI-native product MVPs. Every build ships with an evaluation set, a cost model per interaction and full trace observability — those are not upsells.
Model choice is a routing decision per task, not a brand allegiance: frontier models where judgement quality justifies the cost, smaller and open-weight models where measurement shows no loss. Products are built model-agnostic so a vendor price change is a config edit rather than a rewrite.
Data minimisation first: nothing is sent to a model that the task does not require. Retrieval scopes are tenant-isolated, secrets stay server-side, prompts and traces are retained on a fixed expiry schedule, and training on client data is off by default and never enabled without written consent.
Because the failure mode of this field is confident demos. Publishing the open question, the early finding and the cost of each architecture is how the division stays honest — and it is the same reason every engagement ships with numbers a client can audit rather than adjectives.
006 · COLLABORATION