JP

Intercom Terminal is a terminalcore redesign of Intercom as an interactive AI Support OS. I reimagined the traditional helpdesk website as a live command-center experience where customer conversations become logs, Fin runs like an agent daemon, tickets become queues, and support leaders can operate the workspace through terminal commands. Users can run commands like /scan, /fin, /tickets, /qa, and /briefing to inspect support metrics, Fin AI Agent activity, SLA-risk tickets, quality signals, and support operations summaries. The experience includes command autocomplete, a command palette, live-style support logs, dashboard panels, ticket routing views, Copilot draft previews, QA insights, command sound effects, and a responsive terminal UI. The problem it explores is that modern support teams often have too much information spread across dashboards, inboxes, tickets, reports, and AI agent logs. Intercom Terminal turns that complexity into an operational interface where important signals like SLA risk, topic spikes, sentiment changes, and AI resolution rates are easier to scan and act on. I used v0 to rapidly prototype and iterate on the full website redesign, transforming Intercom’s product structure into a completely different terminal-native visual and interaction system. I used ElevenLabs to create the audio layer of the experience. I designed a custom support-ops voice and generated a spoken manager briefing using ElevenLabs Text to Speech. When the user runs /briefing, the site plays an ElevenLabs-generated support summary covering message volume, Fin resolution rate, SLA-risk tickets, billing topic spikes, Copilot drafts, and recommended actions. The result is a playful but practical vision of what customer support could feel like in the AI Agent era: faster, more operational, more ambient, and easier to understand at a glance. Video - https://www.youtube.com/watch?v=VkC-m-jUOcI demo - https://intercom-clone-omega.vercel.app/

Submitted 7 May 2026

Seven Minutes is an emotional choice-driven narrative game about grief, memory, and the impossible things we wish we could say. You play as Kira, who has spent six years saving for seven minutes with the father she lost. He does not know who she is. She cannot warn him. She cannot change what happens next. All she can do is sit across from him in a diner and choose what to say before time runs out. We used ElevenLabs to create the voice performances, music, ambience, and sound effects that carry the emotional weight of the story. We built the game in Zed, using it to rapidly write, debug, and shape the interactive narrative experience.

Vaidya — Ambient Clinical Scribe for Indian Doctors Indian solo practitioners see 30-40 patients daily. Most write notes on paper or not at all — clinical documentation is the first casualty of a packed waiting room. Western ambient scribes don't handle Hindi-English code-switching, and enterprise pricing is out of reach for small clinics. Vaidya listens to a doctor-patient conversation in Hindi/Hinglish, produces a structured English SOAP note (Subjective, Objective, Assessment, Plan), and generates a patient-friendly Hindi summary — all automatically. **How it works:** 1. Doctor starts a visit → browser captures audio via MediaRecorder 2. Live transcript appears in real-time during the conversation (Scribe v2 Realtime) 3. After "End Visit," the full audio is processed through Scribe v2 Batch with speaker diarization and 196 curated Indian medical keyterms (drug brands, Hindi symptom phrases, Ayurvedic terms) 4. The diarized transcript feeds Google Gemini to generate a structured SOAP note in English 5. A second Gemini call produces an 80-200 word patient-friendly Hindi summary 6. The doctor reviews, edits, and signs the note. The Hindi summary plays aloud via ElevenLabs TTS (Eleven v3) 7. On the patient detail page, a voice assistant (ElevenLabs Agents) lets the doctor ask questions about any patient's history by voice — in Hindi **ElevenLabs integration (5 products):** - **Scribe v2 Batch** — core transcription with 32-speaker diarization, Hindi/English code-switching, and keyterm prompting for medical vocabulary - **Scribe v2 Realtime** — live transcript preview during recording via `@elevenlabs/react` useScribe hook - **TTS Eleven v3** — Hindi patient summary narration with warm, natural voice - **Conversational AI Agents** — voice assistant on patient detail page using React SDK (ConversationProvider + useConversationClientTool for patient context) - **ElevenLabs UI** — 5 components: LiveWaveform (recording visualization), AudioPlayer (visit playback), MicSelector (microphone selection), ShimmeringText (animated branding), Orb (agent speaking/listening state) **Kiro IDE usage:** - 2 specs with full requirements → design → tasks workflow (ambient-scribe-pipeline: 15 requirements, 12 correctness properties; patient-voice-assistant: 4 requirements) - 4 steering docs (externalized LLM prompts for SOAP generation and Hindi summary, medical writing conventions, consent/privacy guidelines) - 2 hooks (typecheck-on-save, test-after-task) - ElevenLabs Power for guided API integration - 75 unit tests, property-based testing with fast-check **Tech stack:** Next.js 15 (App Router), shadcn/ui, SQLite via Drizzle ORM, Google Gemini via Vercel AI SDK, TypeScript throughout. **Pipeline performance:** 18.6 seconds end-to-end (3s transcription + 2.4s SOAP generation + 13.5s Hindi summary) for a 10-minute visit recording.

Underscore turns a filmmaker's creative corpus into a grounded film score — no generic stock music, no guesswork. Upload your scripts, director's notes, subtitles, and moodboards. Every document is parsed, chunked, and embedded using Google's Gemini gemini-embedding-001 (768-dimensional vectors). Both prose chunks and sonic signature cross-embeddings are indexed into Turbopuffer — a serverless vector database — across two per-project namespaces. When you describe a scene, Underscore runs 4 parallel retrieval queries against Turbopuffer: cosine vector search, BM25 full-text search, director-notes-filtered vector search, and a sonic namespace query. Results are fused using Reciprocal Rank Fusion (RRF) to surface the most relevant evidence across all query types. Claude (Anthropic) reads the top retrieved chunks and synthesizes a cue brief — mood, tempo, instrumentation, key themes — grounded entirely in your own materials. It also outputs 3 music prompts (fast burst, cinematic, voice-weighted) and 2 SFX descriptions tuned to complement the scene. ElevenLabs Music (composeDetailed) generates all 3 score variants and a separate 120-second title track cue in parallel. ElevenLabsSound Effects (textToSoundEffects) converts Claude's SFX descriptions into physical, environment-matched audio clips. All generated audio is stored in Vercel Blob (private access) and served via a proxy route. Auto scene extraction uses Claude to identify 3 dramatic moments from the indexed corpus, pre-filling the score generation workflow so filmmakers can start generating immediately after upload. The result: music and sound that actually know your film. PS - The docs/ folder contains a ready-made corpus for the short film The River — use it to test the full pipeline immediately.

VoiceVault helps people facing voice loss preserve their voice in under 5 minutes — and have real conversations with it. Traditional voice banking requires reading 350–1,600 robotic sentences in a clinical setting. Most people give up or never start. VoiceVault replaces that with 7 natural guided prompts — tell us about your favorite meal, share good news, comfort a friend. You sound like yourself because you ARE being yourself. One tap creates an instant voice clone via ElevenLabs Instant Voice Cloning API. The killer feature: Conversation Mode. Every existing voice tool is one-way — you type, it speaks. VoiceVault is the first to close the loop. It LISTENS to the person talking to you using Web Speech API, generates smart response suggestions using AI, and speaks your chosen reply in your cloned voice. It's not just voice preservation — it's conversation preservation. Also includes: one-tap quick phrases for daily and medical needs, smart phrase expansion ("cold" → "Hey, could you grab me a blanket?"), and voice message sharing with loved ones. Built entirely on Replit using ElevenLabs Voice Cloning + TTS APIs. Accessibility-first: WCAG AA colors, 48px+ tap targets, full keyboard/screen reader support. Live demo: voice-vault-cajpany.replit.app https://youtu.be/7Dq4_xUpDkQ

Borderless is a real-time multilingual voice room built for live conversations. One speaker joins a room and speaks once, while listeners open the same room URL, choose their language, and hear live translated audio with matching translated captions. It solves a simple but important problem: multilingual conversations usually break the moment part of the audience cannot follow the spoken language. Borderless turns a single live speaker into a shared multilingual room without requiring separate calls, separate interpreters, or separate TTS generation for every listener. On the ElevenLabs side, Borderless uses: - ElevenLabs Scribe v2 Realtime STT to transcribe the speaker with low latency - ElevenLabs streaming Text-to-Speech with eleven_flash_v2_5 to generate live translated speech for each active language channel On the Cloudflare side, Borderless uses: - Cloudflare Workers for the speaker ingress relay and TTS pipeline - Durable Objects for per-room coordination, ordered sequencing, and language-based fanout - Workers AI for translation before speech synthesis A key design choice is efficiency: Borderless generates one translation and one TTS stream per active language, then fans that stream out to every listener on that language channel. That makes the system both realtime and scalable for live rooms. How to use the demo - Speaker page: https://borderless.jp-45a.workers.dev/room/demo-phase8/speak - Listener page: https://borderless.jp-45a.workers.dev/room/demo-phase8 - Open one speaker tab and one or more listener tabs - On each listener tab, choose a different language and click `Initialize Listener Audio` - On the speaker tab, click `Start Speaker Audio` and begin speaking - For the most stable demo, the speaker should use English input Video Demo Walkthrough - https://youtu.be/Ke043K1dFNw

FlatScout is a voice-first apartment hunting agent that saves renters hours of manual searching. You speak your criteria out loud, and FlatScout searches listings, scrapes property details, verifies neighborhoods, flags red flags like noise or hidden costs, and ranks the best options in a live research dashboard. We built it with ElevenLabs as the core conversational layer: the user talks to an ElevenLabs voice agent, and the agent uses webhook tools to search listings, scrape details, verify neighborhoods, deep-dive into properties, and check live pricing/availability. We also use ElevenLabs client tools to update the UI in real time with listing cards, warnings, search status, activity feed events, and the final comparison table. We used Firecrawl as the web intelligence engine behind the agent. Firecrawl powers city-targeted rental discovery, listing-page scraping, and deeper property research so the agent can move beyond generic search snippets and give structured, decision-ready apartment intelligence. What makes FlatScout special is that it doesn’t just “chat” about rentals. It acts like a real apartment scout: it investigates, verifies, explains its reasoning live, and shows the work visually so you can trust the shortlist. https://youtu.be/bDe4jg4_1IA Username: flatscout Password: 676re6s9cvj0NbeZTOQuab/FNs9LkmjB