# Eight Years of Wanting, Three Months of Building with AI

# Eight Years of Wanting, Three Months of Building with AI

**By Lalit Maganti**
**Published: April 5, 2026**

Source: https://lalitm.com/post/building-syntaqlite-ai/

---

For eight years, Lalit Maganti wanted high-quality developer tools for SQLite. Despite SQLite's ubiquity, the ecosystem lacked reliable, fast, and flexible tooling. Building a proper parser requires extracting SQLite's source code and creating accurate parse trees for roughly 400 grammar rules. This repetitive yet technically demanding work deterred him for years.

## The AI Factor

In early 2025, coding agents improved significantly. He spent January "vibe-coding" — delegating design and implementation to Claude Code on the Max plan — and achieved a working prototype: a C parser extracted from SQLite sources, a formatter, PerfettoSQL support, and a web playground.

However, reviewing the prototype revealed critical flaws. The codebase was "complete spaghetti," with unclear architecture and functions scattered across files. While it proved the approach viable and generated over 500 tests, it required a complete rewrite.

He switched to Rust and completely changed his relationship with AI. Instead of delegating autonomously, he became the decision-maker, using AI as "autocomplete on steroids" with rigorous code review, upfront design, and automatic validation scaffolding. The rewrite took through February, with the final 0.1 release in mid-March.

## Where AI Delivered Value

- Overcoming inertia: transformed abstract uncertainty into concrete prototypes to critique
- Code generation speed: "obvious code" implementing functions matching specified behaviors
- Refactoring at scale: continuous refactoring prevented debt accumulation
- Research and domain bridging: learning Wadler-Lindig pretty-printing, Rust tooling, VS Code APIs in hours instead of days
- Feature completeness: editor extensions, Python bindings, WASM playground, docs, packaging all achievable in timeline

## Where AI Created Problems

- The addiction loop: slot-machine psychology, fatigue makes prompts vague, output degrades
- Loss of mental models: losing touch with which functions call which, accumulated decisions
- Design decision deferral: refactoring felt cheap so architectural choices were postponed
- False test security: 500+ tests created false confidence, prototype still needed rebuilding
- Absent historical context: AI sees present state without feeling how APIs evolved or why decisions were made

## The Core Insight

AI excels where he had deep expertise (parser rules — instant evaluation). Succeeds conditionally with domain bridging (Wadler-Lindig — required engagement). Struggled where he didn't know what he wanted (architecture). Failed critically on tasks without objectively verifiable answers (API design — requires taste).

## The Relativity Framework

Local code (functions, classes) appears Newtonian — AI dominates. Zoomed out, architecture curves unpredictably. Local correctness doesn't guarantee global coherence.

"AI is an incredible force multiplier for implementation, but it's a dangerous substitute for design."