DEV Community: Ritanshu

RELIGHT — AI Cinematic Lighting Engine Built With Kiro | Kiroween Hackathon

Ritanshu — Fri, 05 Dec 2025 16:21:14 +0000

🌟 RELIGHT — Professional AI Lighting Made Accessible to Everyone
Built for the Kiroween Hackathon using Kiro Specs, Hooks, and Steering

Hi everyone! I’m Ritanshu, and this is the story of how I built RELIGHT — an end-to-end AI cinematic lighting engine that transforms ordinary images into studio-quality visuals using a chain of advanced AI models… all orchestrated smoothly with *Kiro.
*
RELIGHT lets anyone create advertisement-level lighting in seconds — without photography studios, complex editing, or professional gear.

And yes…
Kiro didn’t just help me build this, it supercharged the entire development process.
This project simply would not exist at this speed or polish without Kiro.

🎥 What is RELIGHT?

RELIGHT is a multi-model AI pipeline that performs:

✔ Realistic Photometric Relighting (IC-Light) Controls shadows, highlights, mood, and directional light using text prompts.
✔ GPT-4o-mini Vision Lighting Advisor Analyzes the image → suggests 3 optimized lighting prompts under 3 words.
✔ Background-Aware Relighting Matches lighting of the subject with any custom background.
✔ Super Resolution (Real-ESRGAN) Automatically enhances outputs to full 1080p HD clarity.
✔ Clean, Modern UI Drag-and-drop inputs, live preview, light gizmo, and smooth workflow.

*🧠 How RELIGHT Works Internally
*
Here’s the high-level pipeline:

1️⃣ User Upload
The user uploads:

Main subject
(Optional) Background image

Kiro helped me generate the validation code, MIME checks, and async previews.

2️⃣ Background Removal (RMBG 1.4)

The image goes into BRIA RMBG for clean alpha-matte extraction.
Kiro’s vibe coding made it painless to manage GPU memory + batch transforms.

3️⃣ Relighting Stage (IC-Light FC & FBC)

Depending on the mode:

Manual Mode → IC-Light FC
AI Mode → GPT recommends a prompt → IC-Light FC
Background Mode → IC-Light FBC (background-conditioned lighting) This stage injects a lighting embedding into the diffusion process.

Kiro helped me stitch FC + FBC models into a unified class-based architecture.

4️⃣ GPT Vision Lighting Advisor

The uploaded image is encoded to Base64 and sent to GPT-4o-mini.

GPT returns:

✨ “soft top glow”
✨ “left warm beam”
✨ “cinematic right edge”
*This was generated using a Kiro spec.
*
5️⃣ HD Upscaling (Real-ESRGAN x4plus)
Every final relight passes through:
models/RealESRGAN_x4plus.pth

Ensuring:

Zero pixelation
High detail
1080p resolution minimum This step came from a Kiro hook that triggered automatic upscaling after each inference.

🎮 Features That Make RELIGHT Stand Out

✔ Manual Lighting Control
Prompt-based relighting like:

“soft warm left”
“cold top beam” ✔ AI Recommendation Mode
One click → GPT suggests the best moods. ✔ Background Blending Mode
Subject + New Background →
RELIGHT matches:
Light direction
Mood
Color temperature
Exposure
Automatically. ✔ Drag-and-Drop Light Gizmo
A clean UI element that lets users visually move light.

🔧 How Kiro Helped Build RELIGHT (And Why I Love It)

Kiro was not a side tool — it was my development partner.

🔥 Vibe Coding — Instant iteration

I structured all my conversations with Kiro using vibe coding:

“Build me a modular IC-Light wrapper”
“Optimize VRAM for FP16”
“Refactor the upscaling stage into its own class”

The most impressive part?
*Kiro generated a complete relighting pipeline in one go — with model loading, memory handling, and device placement.
*
Crazy efficiency.

🛠 Agent Hooks — Automation that saved my time

I used Kiro hooks to automate:

Running ESRGAN after IC-Light finishes
Clearing CUDA cache between requests
Validating image inputs
Auto-writing spec updates

These hooks acted like a mini CI/CD engine inside my local workflow, making development fast and clean.

📐 Spec-Driven Development — Cleanest architecture ever

I wrote one master spec describing:

Three modes

Inputs/outputs
Expected lighting behavior
File structure

Kiro used this to generate:

Backend structure
API endpoints
Frontend bindings
Documentation

Compared to vibe coding, specs gave me predictable, structured outputs — perfect for a large project.

*🎚 Steering Docs — Sharper, more consistent generation
*

My steering docs made Kiro:
prefer modular code
avoid unnecessary libraries
follow filename conventions
keep GPU memory under 7GB
write clean, readable UI components

The biggest improvement:
Kiro stopped over-hallucinating code, and started generating stable, uniform patterns across all modules.

🎃 Why This Fits Kiroween Perfectly

Kiroween celebrates:

creativity
mixing models
wicked UI
powerful AI workflows
RELIGHT is literally a Frankenstein of:
diffusion
segmentation
super-resolution
GPT lighting analysis
frontend engineering

Plus the UI has smooth transitions, a slick dark-mode theme, and a magical “lighting transformation” effect.

It’s the perfect spooky tech creation.

🚀 Conclusion — RELIGHT + Kiro = Studio Lighting for Everyone

RELIGHT proves that:

💡 Professional lighting can be democratized
⚡ AI can replace expensive photography setups
🎨 Creativity can be automated
👻 And Kiro can turn complex multi-model pipelines into clean, production-ready systems faster than ever

**_Kiro didn’t just improve my workflow…
It elevated it.
It gave me the power to build something that feels like a real product, not just a hackathon prototype.

Thank you for reading — and thank you Kiro for making RELIGHT possible.
Happy Kiroween! 🎃🕯️💡_**

Proj link

🚀 JARVIS — My Journey of Building a Voice-First, Action-Taking Desktop Assistant

Ritanshu — Mon, 15 Sep 2025 18:15:07 +0000

🚀 JARVIS — My Journey of Building a Voice-First, Action-Taking Desktop Assistant

💡 How It All Started

It began with one question:

Why are we still doing boring, repetitive computer tasks the hard way?

I was tired of:

Typing long emails
Jumping between browser tabs
Spending an hour in meetings for two minutes of useful info
Scrolling through endless notifications hoping not to miss something urgent

I imagined an assistant that could do the work, not just answer questions — a voice-first, privacy-focused agent that listens, reasons, and executes.

That idea became JARVIS.

🎯 The Vision

JARVIS isn’t just a chatbot. It’s a voice-first productivity companion that:

Saves hours of repetitive work
Collapses multi-step workflows into one spoken request
Makes computers usable for everyone — including people who can’t rely on a keyboard or mouse

“JARVIS is your AI partner that attends meetings, prioritizes your emails, fixes your code, and handles the busywork — so you can focus on what matters.”

🌟 Hero Features

1️⃣ Attend Meetings & Summarize

Joins Zoom, Meet, or Teams calls automatically
Transcribes conversations in real time with Whisper / Vosk
Extracts speakers, action items, owners, and deadlines
Generates a 30-second briefing so you act immediately

Example Output:

Alice → finish UI design by Thursday

Bob → fix API bug before release

Me → prepare demo slides for Friday

2️⃣ Email Digest & Prioritization

Scans inbox and ranks emails by urgency
Summarizes top messages and drafts replies
Flags deadlines and schedules follow-ups ➡️ No more inbox panic.

3️⃣ Describe My Screen (Accessibility + Dev Tool)

Reads out unread counts and actionable prompts for visually impaired users
Detects error logs and suggests fixes for developers ➡️ Improves accessibility and speeds debugging.

4️⃣ Quick Info & Seamless Browsing

Ask “Jarvis, what is backpropagation?” — get an instant answer without leaving your IDE or video call.

5️⃣ Hands-Free Device Controls

Adjust volume, brightness, or switch apps with a voice command
Control Android via ADB ➡️ Perfect for presentations or multitasking.

6️⃣ Writing & Productivity

Drafts formatted docs, emails, or meeting notes
Auto-names and files them correctly ➡️ Less admin, more creativity.

7️⃣ Smart Communication

Sends Slack/WhatsApp messages, schedules calls, books meetings
Keeps context for seamless follow-ups.

8️⃣ Screen Understanding (Extended)

Takes screenshots → diagnoses issues → pastes fixes ➡️ Eliminates tedious debugging loops.

9️⃣ Learning & Life-Assist

Reads PDFs, recipes, or tutorials step-by-step
Quizzes you hands-free while cooking or coding.

🔟 Time & Device Sync

Voice-controlled reminders, alarms, and events
Cross-device sync with end-to-end encryption.

🔧 Building JARVIS

Architecture

Local-first stack: Ollama LLMs, Whisper & Vosk for ASR
Privacy model: Default local processing, ephemeral transcripts, encrypted sync
Agents: Listener → Reasoner → Action → Accessibility → Connector
Desktop app: Tray daemon + lightweight UI
Plugin model: Meeting → Summary → Create Jira Ticket

Data Flow: Voice → Action

Wake word detection
Audio capture
Transcription (Whisper/Vosk)
Intent parsing (Ollama)
Secure action plan
Execution via OS APIs & connectors
Confirmation + optional logging

⚠️ Challenges Along the Way

Robust speech recognition in noisy rooms
Diarization for multiple speakers
Balancing privacy (local) vs. capability (cloud)
Undo buffer for safe actions
Integrating apps like WhatsApp or Slack
Designing for low-vision users

🏆 Wins I’m Proud Of

Meeting automation → action items in under a minute
Local-first privacy with offline resilience
Accessibility support for blind & motor-impaired users
Modular plugin architecture
Real workflows handled by JARVIS — not just demos!

📚 Lessons Learned

Automation should collapse workflows, not add steps
Privacy by default builds trust
Accessibility improves UX for everyone
Local LLMs shine at intent; cloud excels at heavy reasoning
Clear permissions, logs, and undo features inspire confidence

🔭 What’s Next

Near Term

IDE integration (PR summaries, auto-tests)
Meeting → Jira/GitHub automation
Encrypted sync across devices

Long Term

Proactive assistance & pattern detection
Multi-modal context (webcam + screen)
Voice biometrics for personalization
Enterprise features: RBAC, dashboards, analytics

Business

SaaS + on-prem options
Freemium → Pro → Enterprise tiers
Target markets: accessibility, developer productivity, knowledge work

🛠️ Tech Stack

ASR: Whisper, Vosk
LLMs: Ollama (local), GPT (optional)
Framework: Custom modular agent system
Integrations: Gmail, Calendar, Slack, WhatsApp, IDEs, ADB
Desktop: Electron/Qt HUD + tray daemon
Accessibility: OCR engines, semantic UI parsers
Security: End-to-end encryption, ephemeral transcripts

🌍 Why JARVIS Matters

Turns meetings into clear, actionable notes
Surfaces urgent emails without the clutter
Helps anyone — including people with disabilities — work faster
Speeds up debugging and reduces friction everywhere

🔗 Useful Links

📢 Submission for the Kiro Social Blitz Prize

To enter, I’m posting about my favorite thing about Kiro — how its code generation and hooks supercharged my development workflow — on social media (X/LinkedIn/IG/BlueSky).

I’m tagging @kirodotdev and using the hashtag #hookedonkiro.

📝 Submission for the Kiro Bonus Blog Prize

I’m also submitting this blog post on dev.to/kirodotdev with the hashtag #kiro so others can see how Kiro changed the way I approach development.

💬 Tagging: @kirodotdev

🔖 Hashtags: #kiro #hookedonkiro