Log #004: The Tax Finder Launch -- Multi-LLM Orchestration
• Status: 🚀 Launched / P0 Deployed
• Project: Tax Finder (Old vs. New Regime Comparison)
• AI Collaboration Level: 80% Logic (Claude/Codex) / 20% Integration (Human/Gemini)
• Topic: Moving from "Prompting" to "System Architecture."
TL;DR Launched a privacy-first, zero-data-storage Tax Finder to navigate the April 2026 tax changes. The project proved that the "multi-LLM" approach---using different AI models for their specific strengths is the fastest way to ship.
The Build: The Three-Model Stack
This project was a masterclass in AI-assisted deployment. I didn't stick to one tool; I used a specialized stack:
• The Architect (Claude): I brainstormed the idea with Claude, and it surprisingly generated the entire backend architecture and the initial UI. It understood the "New Rules" of 2026 and the logic for the detailed breakdown without much handholding.
• The Integrator (Gemini Code Assist): I tried to use Gemini to plug the files together but found the speed to be a bottleneck today. Instead of waiting, I took the wheel and handled the deployment manually with Claude's guidance.
• The Polisher (Codex): Once the core was running, I used Codex to refine the specific logic and add the final "beautiful" UI touches.
Human Insight: Privacy & Pain Points
As a PM, I knew the "hook" wasn't just the calculation---it was the trust. By making it open and ensuring it stores zero user data, I'm solving a key psychological barrier for people entering their salary details. The app feels "light" because it doesn't have the baggage of a database.
The Commit
Upgrading to a Claude Pro subscription to streamline the V2 rollout. I need to tighten my deployment pipeline, but for a V1 ship, this is a massive win.
Log #003: The V1 Launch -- Breaking the One-Day Myth
Status: 🚀 Launched | Project: IPL Stat Engine | AI Collab: 50% Logic / 50% Human QA & UI Polish
TL;DR: It took a week, not a day, but the difference between "generated code" and "shipped product" is in the manual quality check.
The Build:
My P0 version became too big. I underestimated how much time it takes to move from "AI can write this" to "This is 100% accurate for every user combination."
Human Insight:
AI is still a "blind" developer. I need to find a way to get UI automatically from sources like Figma or v0 because hand-coding the "beautiful" part is the current bottleneck. But for now, it's live and people are using it.
The Commit:
Gathering user feedback and finding a better design-to-code pipeline.
Log #002: The "Visibility Gap" and the V1 Scope Cut
Status: 🛠️ Refinement | Project: IPL Stat Engine | AI Collab: 60% Logic / 40% Manual UI & QA
TL;DR: The "one-day ship" is a myth for data-heavy products; accuracy and mobile UI are the two front-line battles AI isn't winning yet.
The Build:
The scope for the Stat Engine exploded because "close enough" isn't good enough for sports fans.
• The Logic Trap: AI writes queries in seconds but verifying accuracy across 5+ filters requires hours of manual QA.
• The UI Bottleneck: AI lacks the visual feedback loop I have on localhost. It can't "see" the visual balance on a 6-inch screen.
Human Insight:
I've made the PM call to not ship every engaging idea in V1. The biggest win was the mobile version, precisely because I did the manual work the AI couldn't.
The Commit:
Finalizing data validation for the core query set.
Log #001: The "Dashboard Trap" and the 90% Logic Rule
Status: 🛠️ Building | Project: IPL Stat Engine | AI Collab: 90% Logic / 10% Human
TL;DR: The backend logic was 90% there on the first try but making the data "cool" and mobile-friendly is where the real battle begins.
The Build:
The core logic for a ball-by-ball engine is easy to prompt, but the design is a trap. It is incredibly easy to accidentally build a boring "corporate dashboard" when dealing with sports data.
• Iteration 1: Too "data heavy."
• Iteration 2: Split into Quick Search (query-based filters) and Matchups (Batter vs. Bowler).
Human Insight:
Fitting deep-dive stats into a mobile viewport is a nightmare. I'm currently fighting with a carousel system to keep the data accessible without cluttering the UI.
The Commit:
Finalizing the Kohli vs. Bumrah UI.