DEV Community: AKSHAY P

Building GateUndo: A mobile-first railway gate status app for Kerala

AKSHAY P — Fri, 12 Jun 2026 02:36:22 +0000

How I built GateUndo using Next.js, Supabase, Mapbox, PWA support, and community feedback.

Building GateUndo: A mobile-first railway gate status app for Kerala

GateUndo started from a very local problem in Kerala:

Is the railway gate open, closed, or should I expect a delay?

In many Kerala towns, railway gates are part of everyday travel. A few minutes of uncertainty can change whether someone waits, reroutes, or leaves earlier next time.

So I built GateUndo, a no-login, mobile-first web app for railway gate status, missing gate suggestions, and local gate verification.

Live app:

https://gateundo.vercel.app/

What GateUndo does

The current beta supports:

District-wise railway gate list
Map view using Mapbox
Open / Closed / Unknown community reports
One-tap reporting flow
PWA install support
Missing gate suggestions
Community confirm/wrong voting for suggestions
Admin review before suggestions become live gates
Inactive/obsolete gate state for ROB-replaced or removed crossings
Nearby railway station hints
Conservative safety messaging

The app is currently focused on Kannur first instead of trying to cover all Kerala with weak data.

Tech stack

Frontend:

Next.js 14 App Router
TypeScript
Tailwind CSS
Lucide icons
Mapbox GL JS
PWA manifest + service worker

Backend:

Supabase Postgres
Supabase Row Level Security
Supabase Edge Functions
SQL views for public app data
Admin-only review views

Other tools:

Cloudflare Turnstile for bot protection
PostHog for privacy-conscious analytics
Vercel for deployment

Why I chose no login

For this kind of app, login felt like too much friction.

A user may be sitting near a railway gate with poor mobile data. The app should open fast and allow reporting quickly.

So GateUndo uses anonymous reporting, but with guardrails:

Server-side rate limits
Anonymous device hash
Turnstile checks
GPS proximity validation
Report freshness windows
Admin review for suggested gates

This is not perfect anti-spam, but it keeps the experience simple while reducing obvious abuse.

Status calculation

A basic “latest report wins” model is risky.

So GateUndo uses a more conservative approach:

Recent reports matter more
Old reports decay
If reports conflict too much, the status becomes less confident
If there is no fresh signal, the gate returns to Unknown
Nearby reports are treated as more trustworthy
The app avoids saying anything is “safe”

This matters because a false “open” signal near a railway crossing is not a normal product bug. It can create real-world risk.

Gate suggestions

One important feature is missing gate suggestions.

Users can:

Select a district
Place a marker on the map
Add a road/place name
Optionally add nearby station context
Submit the suggestion

Other users can confirm or mark it wrong.

A suggestion does not automatically become a live gate. It needs enough community signal and then admin approval.

This helps avoid fake or badly placed gates becoming official-looking records.

Inactive gates

Local feedback showed that some old gates may no longer exist because of road overbridges.

Instead of deleting them immediately, GateUndo can mark them as inactive.

Inactive gates:

Show as INACTIVE
Do not accept reports
Return no live status
Stay visible only as local context until reviewed

This is useful when users recognize the place but the crossing is no longer active.

Train movement hints

A few users suggested using train movement data to predict gate closures.

That is a strong idea, but I am not adding automatic prediction yet because I do not currently have a reliable official public live train API that I can safely depend on.

For now, GateUndo supports nearby station hints. For example, a gate can show that train activity near a station may affect it, and users can manually check official railway enquiry sources.

Automatic train activity hints may come later if a reliable data source is validated.

Biggest learning

The biggest lesson so far:

Local utility apps need local truth.

The hard part is not only building the UI. It is verifying whether a gate exists, whether the marker is exactly correct, whether a crossing was replaced by an ROB, and whether users trust the data.

Community feedback has already shaped the product direction.

GateUndo is moving from “live gate status app” toward:

A community-built Kerala railway gate intelligence layer.

That feels more honest and more useful.

What is next

Planned improvements:

Verify more Kannur gates
Add missing gates from local feedback
Improve admin review tools
Add favourite/nearby gates
Improve station hints
Explore route-level delay intelligence later

GateUndo is still beta, but the direction is getting clearer with every piece of local feedback.

App:

https://gateundo.vercel.app/

I Built a Prompt Engineering Suite for Developers — Here's Why Generic AI Prompts Are Killing Your Code Quality

AKSHAY P — Thu, 12 Mar 2026 05:00:58 +0000

I'll be honest with you — I was spending more time writing prompts than actually writing code.

You've been there. You fire up ChatGPT or Claude, type something like "build me a Next.js auth system", and what comes back is either too generic, uses the wrong stack, or confidently generates code for a library you're not even using. Then starts the 45-minute "chat-and-fix" loop that drains your focus and your afternoon.

That frustration is what pushed me to build PromptMint — an AI Prompt Engineering Suite built specifically for developers.

The Core Problem

The quality of AI-generated code is almost entirely determined by the quality of your prompt. Most developers (myself included) don't want to become professional prompt engineers. We want to ship features.

The gap between "what you type" and "what the AI needs to produce great output" is exactly where PromptMint lives.

What is PromptMint?

PromptMint acts as an intelligent intermediary between your raw idea and the AI model. You describe what you want to build, lock in your tech stack, pick your goal mode — and PromptMint transforms it into a structured, production-grade prompt in seconds.

The output isn't just a cleaned-up version of what you typed. It's a fully engineered prompt that enforces your stack, matches the nuances of your target AI model, and follows the CO-STAR framework — an industry-standard structure for high-precision AI instructions.

The CO-STAR Framework

Every prompt PromptMint generates is structured around six layers:

Layer	Role
Context	Project background and environment
Objective	Scoped, specific development goal
Style	Architectural and technical preferences
Tone	Professional, engineer-to-engineer
Audience	Tailored for the target AI model
Response	Multi-layer output — API, Schema, UI

This structure is what separates a prompt that produces boilerplate spaghetti from one that produces something you'd actually commit.

Key Features

🏗️ Tech Stack Enforcement

You select your stack upfront — framework, database, API pattern, styling, auth, language. PromptMint locks it in. The generated prompt enforces these constraints so the AI can't drift into using libraries you didn't choose.

🤖 Conflict Detection

If your idea contradicts your selected stack (e.g., your description mentions Vue but you have React locked in), PromptMint flags and resolves it before generation. No more hallucinated mixed-framework code.

🎯 Adaptive Goal Modes

Nine modes that shift what the prompt prioritizes:

Scaffold — Boilerplate and folder structure, TODOs only
Production-Ready — Error handling, strict typing, edge cases
Refactor — Readability and performance optimization
Debug — Root-cause analysis and regression testing
Performance — Latency and resource efficiency
A11y — WCAG compliance and ARIA patterns
SEO Optimized — Semantic markup and metadata strategy
Micro-Optimizations — Hot-loop and memory tuning
Authentication — Secure identity flows and JWT handling

🎭 AI Model Personas

The prompt's system instruction layer is tailored per model. Claude gets reasoning-first framing. GPT gets precise structural instructions. Cursor AI gets multi-file, IDE-aware context. Each model has different strengths — PromptMint accounts for them.

The Architecture

I wanted the infrastructure to be production-grade without unnecessary complexity. Here's the stack I settled on:

Frontend

Next.js (App Router) — Server components, Server Actions, file-based routing
Tailwind CSS + shadcn/ui — Utility-first styling with accessible, composable components
Framer Motion — Declarative animations with minimal overhead
TypeScript — Strict typing throughout

Backend & AI

Google Gemini API (gemini-flash-latest) — Powers the prompt generation engine. Flash was chosen for its speed-to-quality ratio on structured output tasks.
Supabase — PostgreSQL-backed database, row-level security, and auth. The SDK handles cloud sync for Pro users (history, recipes, usage dashboard).

Deployment & Observability

Vercel — Edge-deployed frontend with global CDN, CI/CD on push, and preview environments per PR.
PostHog — Product analytics and session recordings to understand how developers actually use the tool.
Razorpay — Transaction-based payment processing with zero fixed overhead.

Architecture Decisions

DB Resilience Mode: If Supabase is unreachable, the app gracefully degrades to localStorage — no broken UI, no error pages. Users on the Free plan don't even notice.

URL Length Protection: When 1-click launching a generated prompt into Claude or ChatGPT, long prompts can crash certain browsers via URL length limits. PromptMint handles truncation and encoding automatically before constructing the redirect.

SHA-256 IP Hashing: For guest usage tracking (rate limiting without requiring sign-up), IPs are hashed client-side before being stored. No raw IP data ever touches the database.

Who Is It For?

Indie hackers and solopreneurs — go from "idea" to running boilerplate in minutes, not hours, with a consistent stack every time.

Professional developers — stop rewriting mega-prompts for every feature. Get high-quality code from the AI on the first attempt.

Technical architects — use Engineering Defaults to inject team standards (strict TypeScript, SOLID principles, testing requirements) into every AI-generated output.

Junior developers — use Refactor and Debug modes to see how an expert AI would approach your existing code. It's accelerated mentorship.

What I Learned Building It

Prompt engineering is a UX problem. Developers don't fail at prompting because they're lazy — they fail because there's no structured interface for expressing technical intent. A form with intentional constraints is worth more than a blank text box.

Conflict detection matters more than you'd think. The moment I added automated stack conflict resolution, the quality of generated prompts improved dramatically. Garbage in, garbage out is a real constraint even for frontier models.

Model persona tuning is underrated. The same core prompt performs measurably differently depending on how it's framed for the target model. Spending time on per-model system instructions is not premature optimization — it's the whole game.

Try It

PromptMint is live and free to try — no credit card required.

🔗 promptmint

The Free plan gives you access to core features, 10 goal modes, all model personas, and local history. The Pro plan unlocks the full 50+ tech ecosystem, cloud sync, Engineering Defaults, and unlimited generation.

If you've ever rage-closed a chat window because the AI used the wrong framework for the fifth time — this was built for you.

Drop any questions or feedback in the comments. I read everything.

Built by a developer, for developers. Happy to discuss any architectural decisions in the thread.