🤝 Special Community Acknowledgments
Before breaking down the technical updates, I want to extend a massive thank you to two brilliant engineers here in the community: @unitbuilds and @jeelvankhede.
Thank you both so much for taking the time to thoroughly review my portfolio setup, run technical evaluations on the conversational layers, and show genuine interest in my architectural choices. In today's highly competitive economy, having seasoned developers actively look under the hood of my projects and push me to improve is the absolute best motivation to keep building. I am incredibly grateful for your time and insights!
Building in public is one of the most rewarding parts of being an engineer, but nothing beats the moment another developer scales your ecosystem, pulls apart the layers, and hands you a deep, high-value code audit.
Yesterday, I posted about my interactive portfolio, and a stellar developer here in the community, @unitbuilds, left a brilliant, highly critical breakdown checking my UI/UX, markdown strings, and hardware permission layers.
In today's ultra-competitive tech economy, my absolute priority is pushing myself to learn, adapt, and improve my stack daily to be completely industry-ready. Getting professional criticism from seasoned engineers isn't just helpful—it’s the fastest fuel to sharpen my skills. Instead of just replying, I spent the last few hours writing a structural production-grade patch via Antigravity and deploying it live.
Here is exactly what we refactored based on the audit:
🛠️ The Production Hardening Breakdown
1. The Broken Markdown Chat Render Bug
-
The Issue: The assistant was outputting raw string blocks containing unparsed symbols (
**a-sync**) instead of applying clean styles inside the chat bubble view cards. -
The Fix: I pulled a dedicated markdown parser (
react-markdown) straight into the chat message component logic. The strings now parse seamlessly into clean HTML bolding rules natively, preserving our typography styling without lagging the client loop.
2. Intrusive Microphone Access (UX Flow)
- The Issue: Prompting for hardware camera or mic telemetry immediately upon page load is a jarring user experience.
-
The Fix: I completely overhauled the permission cycle. The platform now mounts with a zero-intrusive sound loop state toggle. Microphone prompts are strictly lazy-loaded—they only fire if a user explicitly clicks
[ CALL LIVE AI AGENT ]or toggles[ VOICE ASSISTANCE ]inside the side-drawer panel. Furthermore, I built a luxury dark hardware disclosure intercept modal that explains precisely what the microphone does before the native browser permission dropdown even appears.
3. Data Privacy Disclosures & Compliance
- The Issue: Visitors are naturally uncomfortable with active mic channels without transparent data containment agreements front and center.
- The Fix: I appended an interactive legal compliance footer directly beneath the Fuel the Code server cards grid layout. Clicking the links now opens beautifully animated, theme-aligned modals tracking data minimization policies (explicitly clarifying that voice operations compile locally inside active browser memory over open-source web-speech pipelines, with zero data recorded or stored on remote databases).
🧠 Deep-Dive: LoRA Weights & ONNX Architectures
@unitbuilds also dropped a fantastic question regarding the underlying machine learning models:
"Did he use a base model and train with a LORA? Or was it trained from scratch?"
To lay out the architecture details: The system runs a multi-stage orchestration pipeline targeting a high-speed, client-side ONNX runtime. It isn't trained entirely from scratch. I built a targeted training dataset of clean vocal attributes and trained a custom LoRA (Low-Rank Adaptation) weight matrix adapter over an open-source topology to keep execution latency low enough for instant response loops. The background automation handles tasks by translating conversational intents into remote shell commands using isolated secure tunnels via a custom serverless proxy routing layer.
🤝 A Massive Thank You to the Community
To @unitbuilds—I am incredibly grateful you took the time to go through my layout configs, stress-test the conversational boundaries, and share your technical perspective. Your feedback pushed this project from a cool concept to a production-hardened web app. I would absolutely love it if you could take another look at the live deployment to see how the updates hold up!
I’d also value your feedback on my broader open-source codebases, especially the main system infrastructure files for Bunny.AI (which I trained utilizing Google Colab blocks, connected via cross-device mobile streams, and compiled down to optimized runtimes). You can review all the raw code repositories directly on my profile:
👉 GitHub: github.com/roshansharma-99
The fresh, hardened build is officially live on my custom domain right now:
👉 Live Portfolio: roshansharma.co.in
I am going to start publishing full video walkthroughs and system code deep-dives of these automation tracks very soon, and I would love to get your continued suggestions, critiques, and insights as I keep expanding this project registry.
Let's keep building! 💻🔥
Top comments (2)
Nawww shucks, thank you for the shoutout, I'll be sure to go through the portfolio again and have a look at your codebases and give you a thorough report back on anything I find that could use some improvement and things that are really well done. I really appreciate the opportunity to go through it again and provide continued feedback.
Great job and keep building!
I gave it a quick run through and mentioned some things to look at to Bunny, hope it helps.