Investigation

Bazaar Intel

A company profile is a database record. A research platform has to reason over it — Bazaar Intel is built as the second thing, not the first.

Category
Product
Status
Shipped, 0→1 build
Role
Product Manager · Founder · Full Stack Builder
Tech stack
Next.js 16, TypeScript, TailwindCSS, shadcn/base-ui, Gemini 2.5 Flash
Scope
120+ companies, 18 industries, 40-50 fields/company
Internal codename
MarketOS (working name during build)
Read time
11 min

Context

Bloomberg, PitchBook, and CB Insights don't exist for Indian consumer internet specifically — the closest equivalents are scattered across paywalled reports, individual company blog posts, and analyst decks that go stale the moment they're published. Anyone trying to build a real picture of a company — its funding history, its competitors, its actual unit economics versus its stated ones — ends up doing that synthesis by hand, every time, from scratch.

The target users were investors, VC analysts, strategy teams, product managers, founder's-office operators, consultants, and startup operators — people who don't need one more company wiki page, they need a research assistant that already knows the landscape and can reason over it on request.

The first instinct was to build better company profile pages. The actual product is a platform that can produce a structured analysis, comparison, or report from that data on demand — the profile page is the substrate, not the product.

Decisions

  1. We assumed

    A market intelligence dashboard's headline numbers — tracked valuation, industry count, public company count, sector leaders — could be set once and updated manually as the dataset grows.

  2. We found

    A hardcoded dashboard drifts from reality the moment a single company is added or a funding round updates, and manual upkeep doesn't scale with a dataset meant to keep growing.

  3. Therefore, we changed

    Built a dynamic analytics engine that computes every dashboard number directly from the underlying dataset: company counts, industry distribution, funding statistics, average valuation, unicorn count, growth leaders. Adding one company updates the entire dashboard automatically, with no manual dashboard maintenance step at all.

  1. We assumed

    Reports (SWOT, risk analysis, growth strategy, competitive positioning) should be pre-written once per company and displayed as static content, the way most research platforms structure their content library.

  2. We found

    Static reports go stale the moment underlying data changes, and pre-writing seven report types across 120+ companies doesn't scale as a one-person build — and more importantly, it caps the platform at whatever reports were anticipated in advance.

  3. Therefore, we changed

    Reports are generated on demand by Gemini 2.5 Flash from the live dataset, formatted as structured markdown and rendered in an executive-report style. The report catalog isn't fixed at build time — a user can ask for a custom analysis type through free-text prompts as well as the six-plus predefined ones.

  1. We assumed

    A single open chat box — 'ask anything about any company' — is the most flexible way to expose an LLM research assistant.

  2. We found

    An unstructured chat box produces inconsistent output shape and makes it hard for a user to know what kinds of questions the tool is actually good at answering, especially for people who want a specific deliverable (a financial snapshot, a risk assessment) rather than a conversation.

  3. Therefore, we changed

    The AI Workspace pairs a structured Company + Analysis Type selector (Company Overview, Competitive Positioning, Financial Snapshot, Strategic Outlook, Risk Assessment, Growth Analysis) with an optional free-text instructions field — giving predictable, professionally formatted output by default, while still allowing an open-ended prompt like 'Analyze Blinkit's profitability compared with Zepto, focusing on unit economics' when a user needs something the presets don't cover.

The product

Bazaar Intel is a seven-page platform: a landing page, a Companies directory, an Industries layer, a side-by-side Compare tool, an AI-generated Reports library, a computed Dashboard, and the AI Workspace — all built around a single structured dataset of Indian consumer internet companies rather than seven disconnected features.

Landing page. Positioned explicitly against Bloomberg/PitchBook-style intelligence tools, scoped to a market none of them cover.
Companies. Category-filterable cards — Flipkart, Amazon India, Meesho, AJIO — each showing YoY growth, funding stage, and headquarters at a glance.
Industries. Each sector carries its own market size, CAGR, and tracked-company count — Marketplace ($85B, 17.2% CAGR, 8 companies) sits alongside Quick Commerce ($6.5B, 45.8% CAGR, 5 companies).
Compare. Any two companies, side by side — industry, founding year, headquarters, employees, valuation, and funding stage, generated from the same structured dataset.

AI Workspace request flow

AI Workspace. Staged progress states — loading profile, reading industry data, evaluating competitors, analyzing business model — instead of a single blank loading spinner during a multi-second generation.
Dashboard. Every number here — tracked valuation, industries covered, public company count, the Market Pulse growth curve, the daily AI insight — is computed from the dataset, not hardcoded.
AI Insight of the Day (live example)"While Flipkart holds the highest valuation at $37B among 120 tracked companies, Quick Commerce signals significant market shifts as the top industry by average growth at 73% across 5 companies, including the fastest-growing Zepto at +150% YoY."

Generated fresh by Gemini from the current dataset state rather than written once and left static — the insight text itself changes as the underlying company and industry numbers change.

Source: Dashboard, AI Insight of the Day panel
Reports. AI-generated research library — Marketplace Landscape 2026, Quick Commerce Report, Consumer AI in India, Fintech Ecosystem — each with a read time and publish date, generated rather than pre-written.

Every company page carries the same 40-50 structured fields: overview, business model, founding year, headquarters, industry, funding, valuation, revenue, employees, market share, growth metrics, profitability, investors, competitors, a SWOT analysis, key products, leadership, strategic initiatives, future outlook, and an AI summary — the same underlying record that powers the Compare tool, the Reports library, and the AI Workspace, so nothing has to be re-entered or re-derived per feature.

Principles

AI gets used where it adds decision-making value — research synthesis, comparisons, risk framing — not layered on as a novelty chat feature. Numbers that can be computed from structured data get computed, never hand-maintained; a dashboard that requires manual updates is a dashboard that will eventually be wrong. And the product is built around workflows a user actually runs — 'what are AJIO's biggest risks', 'compare Zepto vs Blinkit', 'generate a financial snapshot' — rather than around screens that just display information and leave the synthesis to the reader.

Results

00+Companies tracked
00(perscopedocs)Industries covered
40-50Structured fields per company
00+Computed dashboard metrics
00+AI analysis types
00+Pages
00+Reusable React components
Tracked valuation (live dashboard)$358.0B across tracked companies · 20 public companies

Computed live from the dataset, not a fixed headline number — recalculates as company records are added or updated.

Source: Dashboard, Tracked Valuation and Public Companies metric cards

Learnings

Separating the dataset layer from the AI layer, the same discipline behind SellerLens's deterministic scoring engine, mattered here too, just for a different reason: it meant every generated report, comparison, and dashboard number could trace back to a specific field in a specific company record, rather than being an opaque model output nobody could audit or explain.

The AI Workspace's staged loading states (loading profile → reading industry data → evaluating competitors → analyzing business model) came from a simple realization: a multi-second Gemini generation with a blank spinner reads as broken, while the same wait time with visible progress reads as thorough. That's a UX decision, not a performance one, and it changed how the feature felt more than any latency optimization would have.

Notes

  1. Built as a 0→1 solo product build. Company and industry figures shown throughout (valuations, growth rates, funding stages) reflect the platform's structured dataset at time of writing and are illustrative of Indian consumer internet companies, not live real-time financial data.

Related