Loading portfolio
Prem Sai KotaSoftware Engineer
Projects
5 / 11
2026·Local-first financial intelligence·View Website

Unseal

A privacy-first personal finance platform that decrypts and parses bank-statement PDFs inside the app's own server action, with no third-party parsing service, tracking cash burn, recurring subscriptions, and runway.

Next.js
TypeScript
React
Firebase
Groq
Tailwind CSS
Framer Motion
Recharts

Your Bank Statements Do Not Belong in the Cloud

Most personal finance software forces users into an uncomfortable choice: connect bank credentials through third-party data aggregators, or upload raw PDF statements to commercial cloud servers. For anyone who cares about financial privacy, handing over years of transaction history, account balances, and merchant records is an unacceptable risk. I built Unseal to explore a local-first approach to personal wealth tracking. The core premise was simple: statements should be decrypted, normalized and categorized inside Unseal itself — no third-party parsing service — with personal details stripped before any AI sees the data, giving individuals clear financial visibility without exposing private records to external databases.

Parsing Messy Banking PDFs Locally

Bank statement PDFs are notoriously difficult to extract programmatically. Different institutions, such as the State Bank of India, format statement tables with irregular spacing, inconsistent transaction descriptions, and password-protected encryption. I engineered an extraction pipeline using custom regular expressions and text normalization routines. The parser processes PDF byte streams, unlocks password-protected files inside the app's server action, and standardizes transaction rows. To protect sensitive data before optional AI analysis, I built a dedicated PII redactor that replaces emails, IFSC codes, mobile numbers and account numbers with placeholder tags before any text reaches the language model.

Eliminating Floating-Point Drift with Paise Math

Financial software cannot tolerate floating-point rounding errors. In JavaScript, standard IEEE-754 decimal arithmetic produces minute calculation drifts that accumulate across hundreds of transactions, causing ledger totals and summary cards to disagree. While I initially experimented with external currency libraries, they added unnecessary bundle weight. I replaced them with a lightweight native integer engine that computes all balances, expenses, and burn rates strictly in integer paise. Values are converted to decimal rupee representations only at the final presentation layer, guaranteeing mathematical consistency across monthly balance sheets and runway projections.

Hunting Vampire Subscriptions and Runway Bleed

A central goal of Unseal was surfacing recurring charges that quietly drain personal capital—what I call vampire costs. I developed a heuristic subscription detection algorithm that groups transactions by normalized vendor signatures and identifies recurring monthly billing cycles (charges repeating every 28–31 days). The application calculates an adjusted monthly burn rate from real historical data and feeds it into an interactive financial runway gauge. Users can immediately see how many months of operational runway their current liquid savings provide, paired with dynamic spending charts powered by Recharts.

What Unseal Taught Me

Unseal demonstrated that building privacy-first software requires disciplined architecture at every integration point. It is tempting to offload messy parsing jobs to third-party cloud APIs, but doing so compromises the application's core values. Writing unit tests for regex parsers, handling edge cases in malformed statement exports, and writing strict Firestore security rules to protect user-owned documents taught me how to construct defensive client applications. It proved that modern browsers have more than enough compute power to handle complex data extraction privately.