01
Post your image
POST /api/v1/critic/critique with an image URL or data URI, a rubric id, and a reference set id. Your choice of provider (default: OpenAI GPT-4o vision). Soft cap 20MB per image.
Loading…
Insights by Omkar · Studio
Your generator builds. Our critic judges. Against named references — breed standards, brand kits, motion specs — not vibes. Structured critique in, refinement path out. The generator and the critic never share a model. Cross-provider independence is the product.
Stability: beta · SemVer track: critic-v1 · Endpoints live at urja.insightsbyomkar.com/api/v1/critic/*
How it works
The critic never sees your generator code or training data. It receives (a) a rendered image, (b) a named rubric, and (c) a reference set with explicit license + anatomy. It returns structured JSON — per-axis scores, ordered failings with proposed fixes, and a narrative. Your pipeline decides what to apply.
01
POST /api/v1/critic/critique with an image URL or data URI, a rubric id, and a reference set id. Your choice of provider (default: OpenAI GPT-4o vision). Soft cap 20MB per image.
02
The critic consults only the named references. No subject priors, no model memory. Scores every rubric axis 0–10 with mandatory reference citations. Harsh by design — anchor descriptions at 2 / 5 / 8 / 10 prevent model niceness drift.
03
Axis scores, weighted total, top failings (ordered by severity, with landmark ids and proposed fixes), and a 2–4 paragraph narrative. Failings carry autoApplicable flags so you can route numeric deltas to automation and structural work to humans.
04
Apply the deltas your pipeline trusts, re-render, re-critique. Cache keys include provider id so provider swaps never alias old critiques. We ship a ready-made Lucky refinement loop at scripts/critic/refine-lucky.ts in this repo as a reference implementation.
What ships in v1
Rubrics · 2
Seven weighted axes: proportion-fidelity (20%), silhouette-read-at-24px (15%), anatomical-landmarks (20%), line-weight-hierarchy (10%), tonal-plane-fidelity (10%), coat-texture-appropriateness (10%), expression-honesty (15%).
Five weighted axes: palette-coherence (25%), motion-quality (25%), composition-balance (20%), depth-perception (15%), mood-alignment (15%).
Reference sets · 1+
Canonical breed references for Great Dane illustration work. AKC breed-standard text (fair-use critique), FCI #235 key-points, and 14 typed anatomical landmarks with numerical target ranges — muzzle-skull-ratio, body-square-ratio, stop-depth, tuck-up presence, coat-length-short, and more.
The independence rule
This is the only rule we don’t negotiate. A critic that praises its own output is a toy, not an art director. Every critic pass runs on a provider decoupled from your generator’s provider — today that’s OpenAI GPT-4o judging Claude-generated characters. Tomorrow it might be Gemini judging OpenAI, or a committee.
Default critic: OpenAI GPT-4o. Default generator: Claude. Swap either — the cache key includes provider id, so a critique from provider A never masquerades as one from provider B. Audits stay honest.
The critic never sees your generator's source or prompt. It sees the rendered image, the rubric, and the references. Period. Bias from shared training data is the only open-door — and that's why we insist on different providers.
Every axis score must include a reasoning field tying back to a specific reference. Praise without citation is rejected at the parse layer. Shifts LLM niceness-drift into structural failure rather than silent blindness.
The API is versioned (/v1/), the rubric is versioned, the reference set is versioned, the critic engine is versioned. All four pieces are part of the cache key. Upgrade us, your old critiques keep working; downgrade us, your new ones don't alias.
Integration
curl -X POST https://urja.insightsbyomkar.com/api/v1/critic/critique \
-H "x-api-key: $CRITIC_API_KEY" \
-H "content-type: application/json" \
-d '{
"output": { "kind": "image-url", "url": "https://your-cdn/lucky.png" },
"rubricId": "character-figure",
"referenceSetId": "great-dane",
"subjectHint": "Lucky, a fawn Great Dane mascot with black mask"
}'
# Returns: { meta, critique: { axisScores, weightedTotal, failings, narrative },
# cache: { hit, key }, usage: { costUsd, tokens }, provider }Get a key
All three minting paths produce a JWT scoped to the critic track (and/or any other tracks you pick). The raw token is shown once at mint time and never retrievable afterward — copy it, then store it in your secret manager.
admin.insightsbyomkar.com/products/urja — any staff profile (owner / operator / support) picks a scope preset (Critic read-only or Critic use), sets owner + TTL, mints. Raw JWT shown once. Role-based access, TOTP-gated — no shared passwords.
studio.insightsbyomkar.com/account/api-keys — customers view, rotate, and revoke their own keys. Rotation re-mints with identical scopes and emails the new token; the old key is revoked atomically.
POST urja.insightsbyomkar.com/api/internal/provision-key with x-signature: HMAC-SHA256(body, SSO_SHARED_SECRET) and a { nonce, issuedAt } envelope. Body: { owner, plan, ttlDays, scopes }. The Stripe webhook bridge already uses this.
Once minted, export CRITIC_API_KEY=<jwt> and CRITIC_API_URL=https://urja.insightsbyomkar.com. The Lucky refinement loop at scripts/critic/refine-lucky.ts uses exactly these two env variables — no other config.
Pricing
Free tier: rate-limited at the edge. Metered billing comes online with the first paid pilot — per-critique pricing modeled on the provider's underlying vision cost plus a flat orchestration margin. Enterprise: white-label subdomain, custom rubrics, dedicated reference-set curation, SLA. admin@insightsbyomkar.com.