AI Survey Insights India: The AI That Understands Your Survey First

The Survey Knowledge Graph is the technical heart of AI survey insights India on Hercules, and the single reason Poseidon does not hallucinate denominators. At ingestion, Poseidon parses the questionnaire and Parquet schema into four node types. Question Node: one per survey question with id, text, declared type (nps, likert_5pt, multi_select, ranking, text), scale, options, conditional flag and parent condition. Column Node: one per Parquet column with DuckDB type, null rate, cardinality, min/max/mean, top values and statistical profile. Formula Node: one per valid formula with id, valid input types, SQL template and output schema (e.g., {nps_score: FLOAT, range [-100,100]}). Demographic Axis Node: one per demographic question with column name and category levels (city_tier: Tier 1/2/3/Rural). Edges connect them: Question to Column maps_to, Column to Formula formula, Question to Question routed_from for skip logic, and Formula to Demographic Axis slice_by for GROUP BY. Built once at ingestion, traversed for every question.

Why four nodes, not a flat table? Because AI survey insights India must be typed. A multiple_choice stores a categorical string and allows Frequency/Crosstab/Chi-Square; a multi_select stores semicolon-delimited 'A;B;C' and allows Multi-frequency/Co-occurrence; a ranking stores ordered semicolon list and allows Borda count; an nps stores integer 0-10 and allows NPS score; a number_rating stores integer/float and allows mean/median/correlation; a likert_5pt stores scale-bound integer and allows mean/top-2-box; text stores free string and allows Sentiment/Theme; vertical_ranking (grid) stores matrix rows and allows Grid heatmap. Without nodes, a column that looks numeric could be Averaged when it should be Borda-counted. With nodes, AI survey insights India on Poseidon picks the right formula from the column to formula map, not from English keywords, and the LLM is only asked to do what it is good at — language — never arithmetic.

The payoff is visible in answers. Ask 'What is NPS by city tier?' Poseidon traverses: find Question where type=nps, follow maps_to to nps_score Column, follow formula to NPS_FORMULA, find Demographic Axis where label matches city, assemble SELECT city_tier, ROUND((COUNTIF(nps_score>=9)-COUNTIF(nps_score<=6))100.0/COUNT(),1) GROUP BY city_tier, execute via DuckDB, verify re-computed numbers at 99.1%, and return with confidence high. Ask 'most ranked first?' it follows ranking type to Borda formula automatically. Ask 'skip-logic churn reasons?' it follows routed_from edge to add the correct WHERE. That typed traversal is why AI survey insights India on Hercules works across 8 plus languages without translating the questionnaire; the graph knows structure independent of language, and built by Jupiter Meta Labs in Bangalore, it is the moat that turns structure into insight before a word of narrative is written.

Column to formula mapping is where AI survey insights India becomes formula-driven, not prompt-driven — the opposite of generic chat tools that ask an LLM to 'do some math.' The mapping table is explicit: multiple_choice → Frequency/Crosstab/Chi-Square; multi_select/checkbox → Multi-frequency/Co-occurrence; ranking → Borda count/Median rank/Top-2; nps → NPS score/Segment split; number_rating → Mean/Median/Distribution/Correlation; likert_5pt/7pt → Mean/Top-2/Top-3/Distribution; text → Sentiment/Theme/Keyword; vertical_ranking grid → Grid heatmap/Row average. When you ask 'average satisfaction for delivery,' Column Selector finds the likert_5pt Node, follows its formula edges, and the Code Generator receives the valid options — mean and top-2-box — with the SQL template for each. No LLM invents the formula; the graph supplies it, the generator assembles it, DuckDB executes it, and the verifier re-derives it. That chain is how AI survey insights India earns the word verified.

Confidence tags on edges add auditability that generic AI lacks. EXTRACTED means the relationship was directly declared — e.g., question type is nps, so NPS edge is certain. INFERRED means fuzzy match resolved it with a score — e.g., 'preferred payment mode' matched a multi_select column with confidence 0.72 — reviewable and often correct but flagged. AMBIGUOUS means multiple formulas could apply — e.g., a numeric column could be scale or raw count — so Poseidon asks for clarification rather than guessing. Those tags flow into reporting as confidence low/medium/high shown in chat, and into quick suggestions as two to three clickable next analyses that are guaranteed valid for the current graph. For a Mumbai team asking 'what drives NPS?' the suggestions are not guesswork; they are Formula to Demographic Axis slice_by edges that are known to be valid GROUP BYs for NPS. AI survey insights India on Hercules thus teaches you what to ask next, with confidence you can act on.

The table also prevents method mismatch that is costly in India. A founder in Indore asking 'which feature is most important?' could mean MaxDiff for forced choice or Borda for ranking — the graph checks questionnaire type and routes to Borda if the column is ranking, or surfaces both with clarification if AMBIGUOUS. A Hyderabad pricing ask 'too cheap versus expensive' maps to Van Westendorp only if the four price columns exist with correct types; otherwise Poseidon explains what is missing rather than inventing a price curve. A Chennai 'brand funnel' maps to funnel conversion only for brand tracking survey types with awareness→recall→consideration→preference→loyalty questions. Each mapping carries a SQL template with output schema checks — e.g., NPS output range [-100,100] — enforced by Output Validator in lite or full mode. That is AI survey insights India that is correct by construction, delivered at Free ₹0/month to start and Pro ₹30,000/quarter for depth, all via Poseidon on hercules.works/ai.

Denominator correction for skip logic is the most Indian of AI survey insights India problems — and the graph solves it structurally. Most surveys have conditional questions: Q12 'Why did you churn?' asked only if Q3 equals churned or NPS ≤6. Raw data therefore has nulls in Q12 for non-churners, and any naive analysis dividing by 400 overstates conditional insights. Poseidon detects routing pairs at ingestion via Survey Intelligence that scans null-correlation in Parquet: if Q12 is non-null only when Q3 is churned, it creates a routed_from edge carrying the showIf/next rule. Then, when you ask 'What are churn reasons?' the Column Selector finds Q12, follows routed_from to the parent condition, and the Code Generator automatically adds WHERE Q3 equals churned — so the denominator is 112 churners, not 400 total. Output Validator checks that n in the narrative matches that conditional base, and Self-Critique warns if the narrative omits it. That auto-correction is why the Pune founder's 68% versus 26% error cannot happen on Hercules.

Multi-select normalisation is the companion fix. Raw multi_select is stored as one row per respondent with semicolon-delimited 'Feature A;Feature B;Feature C.' At query time, DuckDB's UNNEST(string_split(col, ';')) expands each row into one option row, so COUNT(*) counts selections, not respondents, and percentages can exceed 100 correctly for multi-frequency. Ranking is stored as ordered semicolon list and expanded similarly for Borda count via Σ(n − rank_position). Without the graph knowing the column type, AI survey insights India would conflate respondents and selections and mis-rank features. With it, a Kochi D2C question 'which features are most selected overall?' correctly reports selections with note 'denominator is selections,' while 'how many respondents selected feature A?' reports respondents. Both answers show base N and are three-layer verified at 99.1% with 78% cache hit, so the second identical question returns under 200 ms at zero LLM cost.

Together, these two fixes make AI survey insights India trustworthy at decision time. A Bangalore edtech asking 'what pricing would premium seekers tolerate for AMOLED?' gets a Gabor-Granger curve where conditional base 'premium seekers' is correctly defined by segmentation, with WHERE applied from the routing edge that defines the segment battery, and with Van Westendorp curves per language where Hindi and Tamil are separate slices via Demographic Axis. A Government of Karnataka citizen survey asking 'satisfaction among ward respondents who experienced water shortage?' gets the correct ward-level denominator, not city total. Built by Jupiter Meta Labs in Bangalore with type-aware analysis, ephemeral DuckDB connections and AST sanitisation, AI survey insights India on Hercules ensures that if a slice cannot be correctly derived, it is flagged or clarified — never shipped with the wrong base, and that certainty is available from Free ₹0/month with 10 AI chats through Starter ₹1,119/month to Pro ₹30,000/quarter.

Quick suggestions and confidence scores are the interface layer that turns AI survey insights India from an answer into a habit. After each verified answer, Poseidon assigns a confidence score low/medium/high based on verification passes, sample size and edge confidence (EXTRACTED versus INFERRED versus AMBIGUOUS), shown in the chat beside verification metadata. A 'mean rating for delivery speed in Tier 2 (n=412, high)' inspires different action than 'price sensitivity among Tier 3 premium seekers (n=42, low — small base, interpret cautiously).' That calibrated humility — built into the 5-phase Reporting phase after Verification — is itself a trust moat: the system tells you when to bet and when to re-test. For a Delhi CMO who must decide today with 42 Tier 3 premium seekers versus waiting a week to fill the quota, confidence plus n is the input to a rational call.

Quick suggestions are two to three clickable chips under the answer, generated from the Survey Knowledge Graph's Formula to Demographic Axis slice_by edges and from high-leverage derived tasks added by Analytics Planner (weighted scores, funnels, driver Shapley, importance×performance). After 'NPS by city tier,' you might see 'split NPS by top-ranked feature,' 'show top-box drivers for Tier 2,' or 'Van Westendorp curve for premium seekers' — each guaranteed valid for your questionnaire, not a generic 'try a bar chart.' Clicking one reuses the 5-phase pipeline, hits column profiles from cache, often hits semantic cache at 78% for instant under-200 ms return, and triggers the same three-layer verification. That is AI survey insights India that compounds: each answer teaches you the next best question, so a growth lead in Pune who knew what she wanted but not the method name ('should I do MaxDiff or ranking?') is guided to the right method via the graph's valid edges without learning terminology.

Together, suggestions and confidence convert a one-off chat into a continuous insight habit that SuperJ's scale sustains. With 20M plus verified Indians at 60-90%+ rates across 500 plus cities and 8 plus languages, the graph grows with each wave, and the suggestions grow smarter, pointing you to cohort and segment insights that span your tracker. Streaming SSE with 0-100% progress and persistence ensures that even when a suggestion fans to a COMPLEX query, you see progress and can recover after refresh via chat_turns.stream_logs. All at pricing that encourages habit: Free ₹0/month permanent for 10 AI chats and 100 users, Starter ₹1,119/month (₹895 annual) for 2-4 studies monthly, and Pro ₹30,000/quarter (₹24,000 annual) for weekly pulses, all with 100 free responses month one. That is AI survey insights India on Hercules Works — graph-understood, formula-driven, confidence-scored, suggestion-guided and cache-fast, trusted by Unilever through SBI Mutual Fund.

Putting it all together, AI survey insights India on Hercules Works follows one disciplined order: ingest and understand, then ask and interpret. Responses from SuperJ arrive as JSON, are typed via Question Nodes, normalised and stored as CSV then Parquet at data/{chat_id}.parquet, and the Survey Knowledge Graph is built with Question, Column, Formula and Demographic Axis nodes plus maps_to, formula, routed_from and slice_by edges, with confidence tags. Then your English question — 'what is NPS by city tier?' or 'which concept wins on purchase intent among Tier 2 mothers?' — is checked for relevance, rewritten for precision, matched at 0.88 to semantic cache at 78% hit, routed by intent and SIMPLE/MODERATE/COMPLEX, analysed by parallel sub-agents profiling columns and selecting columns via the graph, verified by deterministic Numerical Claim Verifier at 99.1% plus structural Output Validator plus 0-10 Self-Critique enforcing 'So What,' and reported with confidence and quick suggestions. Only after the graph is traversed does the LLM assemble SQL from column plus formula template; it never invents arithmetic. That is the single sentence that captures Poseidon's philosophy: other tools ask the LLM to figure out the data; Poseidon understands the data first, then asks the LLM to do language.

The business consequence is that a non-technical brand manager in Lucknow gets the analytical depth of a senior research analyst in hours, not weeks, and at a fraction of legacy cost — 10-100× cheaper than ₹10-50 lakh agency decks — while an enterprise insights head in Mumbai gets verification and audit trails that satisfy risk and compliance. Whether you run Brand Tracking with awareness→loyalty funnel, Customer Profiling with cluster heatmaps, CSAT/NPS with promoter/passive/detractor cuts, Product Feedback with top-box per feature, Feature Prioritisation with Borda/MaxDiff, Concept Testing with purchase intent, Ad Effectiveness with aided/unaided recall, Pricing Research with Van Westendorp/Gabor-Granger, or Qualitative theme clustering — 12 plus typed templates — Poseidon picks the method that fits the column type, not the English phrasing, and proves the number before you quote it. All charts use colourblind-aware Okabe-Ito palette, Inter font and direct labelling, embedded via chart tokens {{CHART_task_1_overall}} in reports, so AI survey insights India ends with a visual you can send.

Start with the question you have been postponing because 'data is messy.' Hercules Works cleans the mess structurally: correct UNNEST for multi-select, correct Borda for ranking, correct NPS formula, correct WHERE for skip logic, and correct slice for demographics, all via the Survey Knowledge Graph that other tools lack. Type your question in English, watch streaming progress, see confidence and verification badge, click a suggestion to go deeper, and generate a 20-50 page Markdown/HTML/PDF report when ready — all on hercules.works/ai, powered by Poseidon (FastAPI, LangGraph, Gemini, DuckDB, Parquet), anchored on SuperJ's 20M plus ZK-verified Indians, and priced at Free ₹0/month permanent, Starter ₹1,119/month (₹895 annual, 20% off) and Pro ₹30,000/quarter (₹24,000 annual, 20% off) with 100 free responses month one. AI survey insights India that understands survey structure first is not a feature; it is the foundation for every verified insight, quick suggestion and confident decision you will make next.