Everything I've
shipped & studied.

Why

Founders burn hours in spreadsheets before every board meeting — and get different answers each time. The ritual becomes one import button.

How

Transaction-level ingestion, a movement-classification engine assigning every dollar to new/expansion/churn/retained/reactivated, and views shaped like investor questions.

Learnings

• Normalization IS the product — every export disagrees about refunds.
• Classification must be explainable per-row or finance won't trust it.

Why

Revenue data is radioactive — teams hesitate to upload it anywhere. Privacy by architecture: all parsing and math run client-side.

How

TypeScript + custom chart library (Sankey, waterfall, cohort heatmaps), auto-detected remappable columns, data-issues fixer pipeline.

Learnings

• "Never leaves the browser" closes more users than any feature.
• Client-side cleaning UX is the hard part, not parsing.

Why

Replaces the lead spreadsheet every early team runs on — and does the founder math a spreadsheet never will.

How

Stage probabilities roll card values into weighted pipeline and projected MRR; spend entry feeds blended CaC and channel CPL/ROAS. Metrics layer unit-tested.

Learnings

• Fewer fields with the math done beats more fields.
• Auto-logging is the difference between a CRM and a graveyard.

Why

Parsers — AI ones included — break on large multi-page statements and fail silently. In financial data, silent failure is the worst kind.

How

Format-aware parsing per Indian bank layout; a verification layer treats the statement's own totals as ground truth — opening + transactions must equal closing.

Learnings

• Trust the document's totals, never the parser.
• Boring per-format parsers beat clever generic ones 10×.

Why

Only a handful of people understand how the tables relate; onboarding runs on tribal knowledge and Slack archaeology.

How

Explicit FKs first, then inference from naming conventions (user_id → users); interactive graph with per-table focus; AI explains what clusters are for.

Learnings

• Conventions encode 80% of undeclared relationships.
• Progressive disclosure is the real design problem.

Why

Indian SMBs juggle GST compliance, Tally and collection as three disconnected chores; money falls through the cracks between them.

How

One GST engine drives preview and server so they can never disagree; Tally XML with queued retries; bank credits auto-match with confidence tiers — high posts automatically, the rest queue for review.

Learnings

• Money math in floats is a bug factory — integer paise everywhere.
• Queue-and-retry beats "is Tally running?".

Why

Bookkeepers spend hours categorizing transactions QuickBooks has seen a hundred times before.

How

Exact-vendor, fuzzy and recurring-payment matching against the client's own history; settles invoices/bills; detects transfers instead of double-counting. Six test modules.

Learnings

• Human-in-the-loop is the feature accounting buyers ask for first.
• Transfers are where naive categorizers double-count.

Why

Built to run a real agency end to end — marketing, conversion, collection and operations in one place.

How

React 19 + GSAP 3D scroll homepage, dual layout system, custom prerendering build pipeline, Express 5 + Neon backend, better-auth OTP, Razorpay webhooks.

Learnings

• Prerendering beats SSR for weekly-changing content.
• Source attribution must be captured at write time.

Why

Product, growth and revenue data lived in four tools; every question needed four tabs and manual joins.

How

Identity-joined pipeline across product DB, PostHog, Stripe and Meta; drop-off detection fires re-engagement notifications automatically.

Learnings

• Four sources is easy; four teams agreeing on definitions is the project.
• Metrics wired to actions change behavior; dashboards alone don't.

Why

App-store reviews were a goldmine of product feedback nobody was reading systematically.

Learnings

• Reviews cluster into few themes fast; the long tail is thin.
• NPS by version turns "users are unhappy" into "release X broke Y".

Why

To prove BI tools can do more than bar charts — recommendation logic with pure LOD expressions and parameter actions.

Learnings

• Tableau computes almost anything if you think in LODs.
• Parameter actions are a full interaction model.

Why

Everyone who commutes by train or bus has slept past their stop.

How

MapLibre + OpenFreeMap + Photon + OSRM (no Google billing); unit-tested pure-Dart trip state machine; alarm on the alarm audio stream; distance-based snooze; CI APK builds.

Learnings

• Free map stack works; geocoding quality varies outside big cities.
• Battery-vs-accuracy tiers beat sliders.