Survey Deployment on SuperJ in India: From questions_json to Poseidon Parquet

What Actually Leaves Your Builder — questions_json, Not a Form Link: On Hercules Works, every survey ready for field exists as questions_json versioned in survey_chats and survey_versions — a typed array of questions with stable qId and human-readable id, types like multiple_choice, ranking or vertical_ranking, options, media refs and routing compiled to next, showIf, terminate and complete. When you deploy, the backend does not email a link. It converts that internal schema to SuperJ’s Mongo payload via the SuperJ service and registers or updates the survey shell through SUPERJ_CREATE_SURVEY_AI_URL, the managed endpoint that creates the SuperJ shell and maps audience_payload alongside questions in one atomic push. That mapping is why a ranking, a grid or a text open-end renders correctly on the SuperJ app in Hindi, Tamil or Hinglish without you rebuilding anything for mobile. Built by Jupiter Meta Labs in Hyderabad, this separation keeps authoring in Hercules and execution on SuperJ clean. Poseidon later relies on that same typed schema to build the Survey Knowledge Graph once at ingestion and traverse it for every query, while semantic cache at 78% hit rate keeps repeat deployments of trackers fast and cheap.

Versioning, Previews and Media Placeholders in the Same Push: Questions of type image_multiple_choice or video ship with a placeholder URL that is later injected via POST /api/chats/{id}/inject-media when you upload an asset; the backend attaches it to the correct qId so the SuperJ renderer shows the right 1080 image or non-skippable video frame without breaking routing. Stimuli are tagged to stimulus IDs so rotation and recall are tracked deterministically per respondent, which matters for SD-036 concept cards where order bias must be controlled. Every deployment is versioned — you preview any version before going live via GET /api/chats/{id}/versions/{n}/preview — so a last-minute Hindi correction to a pack claim does not overwrite history or orphan prior completes. See how payloads link to verification at survey data verification India and how the engine understands structure via survey knowledge graph. On Hercules Works this is painless because the SuperJ panel and Poseidon handle wiring while you keep full control over audience, language and logic.

Why SuperJ App Delivery Beats Email Routers on Reach and Speed: Because fielding happens on the SuperJ app where people answer surveys in exchange for rewards, average completion sits at 60-90%+ versus 10-20% for emailed Western panels that blast English links, and you reach 20M+ ZK-verified Indians across Tier 1, Tier 2 and Tier 3 without renting a separate panel or buying sample blind. SuperJ is the survey app, not WhatsApp — delivery is inside the app with notifications, rewards accounting and ZK deduplication, not a chat thread. That app-native design is why a 400-sample study closes in hours not weeks, why video forced exposure actually forces, and why Poseidon then ingests a clean Parquet with correct denominators and exposure flags row for row. Compare reach at consumer panel India and the analytics handoff at poseidon analytics engine, both built for India first.

Demography Lives in audience_payload — Never as Survey Questions: One rule Hercules enforces saves every deployment from inflated LOI and broken logic: age, gender, city tier and NCCS never appear as questions. The brief step captures an audience_payload like {"NCCS": {"A1": 50, "A2": 100}, "cities": ["Mumbai", "Indore"], "age": {"18-24": 75}, "total": 400, "checkCity": true} alongside title and created_by, and that payload is what SuperJ uses to sample before the first question is shown. The LLM that drafts questions is explicitly forbidden from inventing demographic questions, so the questionnaire stays focused on behaviour, attitudes and drivers while quotas and eligibility stay in the audience system where they belong. That separation avoids the classic Indian-panel error where screening inside the survey inflates length, causes drop-off, and breaks skip logic because a screened respondent already saw half the survey. Learn routing that respects this at skip logic survey tool and method framing at research methodology India.

NCCS, City Tiers and Age Buckets Built for Bharat: Targeting understands NCCS A1 through C1, Tier 1 metros Mumbai, Delhi, Hyderabad, Pune, Chennai, Kolkata and Bangalore, Tier 2 hubs Lucknow, Indore, Kochi, Jaipur and Ahmedabad, and age buckets 18-24, 24-35, 35-45 and 45-55, with general-population mode setting checkCity false for broad India reach when you brief “All India”. You say “NCCS B1/B2 women 25-34 in metros and Tier 2, 400 completes” and the payload is built deterministically, not by an interviewer’s gut. This precision is why a pricing study can balance premium seekers in South metros without starving the North Tier 2 cell, and why Poseidon’s later Survey Knowledge Graph has clean Demographic Axis nodes for GROUP BY without extra cleaning or recoding. Scope similar targeting depth at tier 2 tier 3 consumer research India and panel scale at superj earn money surveys India. Built by Jupiter Meta Labs in Hyderabad, the audience system reflects how India actually consumes — by class and city tier together.

What Clean Separation Unlocks for Quotas and Analysis: Because audience_payload is a separate contract, Hard, Interlocking, Soft and Dynamic quotas — SD-026 through SD-029 — are enforced at the sampling gate, and Poseidon later reads the same payload as one of five report context sources alongside research brief, questionnaire, Parquet responses and auto-discovered routing pairs. That means denominator correction for conditional questions and correct base sizes in reports without you hand-coding WHERE clauses, and goal narratives that frame findings around the intended who, not just the realised who. When the survey completes and an admin marks it completed, Poseidon pulls audience_payload directly from Postgres and explains coverage. If you landed here from quota questions, the full gate mechanics are at survey quota management India; if you came from analytics, see quantitative research methods India.

Concept Cards and Pack Shots at 1080, Tagged to Stimulus IDs (SD-036): Most Indian buys are decided by a 3-second pack read on a crowded shelf, so SD-036 Concept Card / Pack Shot Display requires controlled image display — zoomable, timed, mobile-responsive — tagged to stimulus IDs so rotation tracking is deterministic. On Hercules, an image_multiple_choice or image_question with image 1080 travels inside questions_json, its URL injected later, and its ID is the key the rotation and recall questions reference. SD-039 Forced Exposure Timer can hold that pack for a mandatory minimum view time before the respondent can proceed, preventing the cursory swipe that corrupts concept and pack data in image-heavy tests for FMCG. Explore visual testing context at ad effectiveness research India and product concept testing platform India. On the SuperJ app this renders natively, rewards sustain 60-90%+ completion even with image stimuli, and exposure flags land in Parquet row for row.

Video Forced Exposure and Comprehension Logic (SD-038): For ads, SD-038 Video Stimulus Insertion is non-skippable playback with comprehension and recall questions triggered automatically upon completion. In deployment, that is a video type question with forced completion flags in the SuperJ payload; the next question’s showIf waits for video completion, not a self-reported “I watched it”. That guarantee matters when you test a TVC in Hindi and Hinglish across metros — you know every claimed recall came from someone who actually saw the film, not a skimmer. Poseidon later treats video exposure as a routing pair for denominator correction so your recall metric is not inflated by non-viewers, and the Survey Knowledge Graph models the exposure edge explicitly for later cross-study linkage. Compare media handling at multilingual survey tool India and data-quality safeguards at survey data quality India. Built by Jupiter Meta Labs in Hyderabad, video is app-native not link-out.

Why Stimuli Must Ship With the Survey Shell, Not After: When stimuli are managed outside the survey shell, rotation logs are lost, recall questions misfire, and Parquet arrives without exposure flags. Hercules ships questions, routing, media refs and audience in a single push, so the Parquet that lands for Poseidon has exposure aligned per respondent for deterministic joins. That alignment is why the 18-node report can merge media diagnostics with preference and pricing without manual VLOOKUPs, and why chart advisor picks the right visual for stimulus-driven lifts. For the full catalogue of stimuli — concept cards, shelves, video, timers, card sort, drag-rank and swipe — see survey stimulus management India, which maps SD-036 through SD-042 to SuperJ mobile execution. The payoff is faster trust: media that was guaranteed seen produces recall you can defend in a boardroom.

Hard Quotas (SD-026) Enforced Pre-Survey at the Sampling Gate: Quotas on Hercules run before an answer counts, via the audience system that SuperJ enforces while sampling from 20M+ verified Indians. SD-026 Hard Quotas are absolute ceilings per segment — for example NCCS A1=50, gender 50/50, Mumbai=40, Hyderabad=30 — where over-quota respondents are terminated immediately on detection, not after they finish a 20-minute LOI. That protects representativeness and panel experience on SuperJ, where a terminated user is routed respectfully rather than wasted. Contrast this with questionnaire screening where you pay for stranded completes that still consumed panel attention; gate enforcement keeps cost per verified complete low across Free, Starter and Pro plans and keeps field honest. See full quota mechanics at survey quota management India and broader quality context at best practices for improving data quality in online surveys.

Interlocking Cells Define Intersections, Not Marginals (SD-027): SD-027 Interlocking Quotas define limits on intersections like Female AND 25-34 AND Metro, not isolated marginals. That prevents the fielding failure where you hit gender overall but miss young metro women entirely — a fatal skew for beauty, quick-commerce or premium mobile accessories where cohort matters. On Hercules this is expressed inside audience_payload as nested quotas that SuperJ samples jointly; Poseidon later surfaces demographic coverage alongside the research brief so the report flags under-filled interlocks instead of hiding them behind topline totals. For shopper examples that need aisle balance see market segmentation survey India and for rural spread see rural consumer research India. Built by Jupiter Meta Labs in Hyderabad, this interlock design reflects how Indian consumption segments by city tier and class together, not one axis at a time.

Soft Targets and Why You Still Need SD-028 Alongside Hard Ceilings: Not every niche can carry a hard ceiling without stalling field — SD-028 Soft Quotas set preferred targets that guide routing and sampling priority without hard terminates, useful for low-incidence behaviours like recent premium TV buyers. The deployment remains smooth because field does not stall waiting for the last 5 Niche-B2 completes; you get a representative core while niche shortfalls are flagged for weighting or a follow-up wave rather than blocking the study. When you pair this with SD-029 Dynamic Rebalancing, field adjusts sampling toward under-filled cells in real time without manual re-uploads or WhatsApp follow-ups. On Hercules Works this is painless because the SuperJ panel and Poseidon handle complexity while you keep control — and your deployment finishes in hours, not weeks, with verified humans only via ZK-verification and zero bots at every gate.

Admin PATCH completed — The Fire-and-Forget That Opens Analytics: When field reaches total, an admin sets survey_status to completed via PATCH /api/update-status or the admin review endpoint POST /api/admin/review-content/chat/{chat_id}; the backend writes survey_status completed, creates a system chat_turns row of type ANALYTICS, and asynchronously POSTs to POSEIDON_INGEST_URL/api/v1/ingest/ with {chat_id, super_j_survey_id} and header X-Secret-Key: <JWT_DRAGON_SECRET>. It returns 200 OK immediately — fire-and-forget. This is the only moment field becomes insight. Poseidon then reads business_flow, routing_logic, audience_payload, formulated_goals and the deployed questions_json directly from PostgreSQL, fetches raw answer data from SuperJ, cleans it into a columnar Parquet dataset, and writes back parquet_path, analytics_data_status ready, analytics_row_count and analytics_column_count. If Poseidon times out after 120s or returns non-200, completed stays and analytics_data_status goes failed for ops to retry without losing field. See architecture at poseidon analytics engine and verification layers at survey data verification India.

Parquet, Not CSV, and Why It Matters for Every Question You Will Ask Next: Parquet is columnar; DuckDB reads only the columns a query needs, skipping the rest. For a 50,000-response study with 40 columns, a two-column aggregation reads under 5% of the file — a 10-20× speed gain over row-based CSV, with dramatically lower memory because no full RAM load is needed. Poseidon runs DuckDB in-process with ephemeral connections, no shared state and no cross-survey leakage, pushing filters at storage. Median per-query time is about 1.8s, cache hits under 200ms, with semantic cache at 78% hit rate and automatic invalidation when Parquet schema hash or row-count fingerprint changes, plus Lite verification mode for low-stakes queries. That stack is why “what is NPS by city tier?” returns a verified chart and narrative in seconds, not a pivot you build. Dive deeper at quantitative research methods India and market research report automation India.

ANALYTICS Turn, 5 Phases and Streaming That Survives Refresh: Ingestion creates an ANALYTICS chat turn; any natural-language question then hits POST /api/v1/analysis/query/{chat_id} through Poseidon’s 5-phase pipeline — Initialization, Routing, Analysis, Verification, Reporting — backed by LangGraph and Gemini, with three-layer verification at 99.1% (deterministic claim verifier that re-runs DuckDB SQL, structural validator for ranges and sum-to-100, and LLM self-critique for narrative quality). Streaming uses Server-Sent Events with 0-100% progress, and if the browser refreshes the backend job continues; result is persisted to chat_turns.stream_logs and recovered by chat_turn_id, so no work is lost. Alongside chat, Poseidon can generate a 20-50 page Markdown/HTML/PDF report via its 18-node graph that writes around the research brief with a “So What” rule — business implication before evidence. On Hercules Works by Jupiter Meta Labs in Hyderabad, this lifecycle — questions_json → SuperJ → PATCH completed → ingest → Parquet → ANALYTICS — closes the loop from idea to board-ready insight within hours at subscription pricing from ₹0/month, not agency weeks.