Team
MirrorMatch
Project Concept
No description has been added yet.
Entry
Status: Submitted
Last saved: June 06 at 7:55 PM EDT
Team Roster
You must be registered for the event to view the team message board.
Advik Bhatt Team Lead RSVP Approved
Founder at Rolemate
Advik Bhatt — Solo project. All code written during the hackathon.
Built the complete full-stack platform in a single day.
Tavus: Integrated CVI via POST /v2/conversations to embed a live AI avatar as the customer service agent.
ElevenLabs: Integrated TTS (eleven_multilingual_v2) with emotion-modulated voice stability. When escalation rises, the Adam voice speaks the coaching action aloud — stability drops as anger increases so the delivery gets more urgent.
Veris: Integrated via the HTTP channel. veris.yaml defines /api/simulate as the evaluation endpoint — Veris sends scripted customer utterances turn by turn, MirrorMatch classifies emotion and returns the coaching signal for that escalation level. Simulated turns are logged to Redis alongside live call data. Dockerfile.sandbox included for containerized deployment.
Redis (Upstash): Built the real-time data layer — emotion arc as a timestamp-scored sorted set, peak-anger ZREVRANGE lookup, atomic INCR counters for turns and escalations, live emotion-state hash read back into the dashboard's in-call REDIS stats chip. Also wired per-IP rate limiting using the same atomic counter pattern: 40 HuggingFace calls per minute and 3 Tavus sessions per minute per IP. 1-hour TTL on all session keys.
HuggingFace: /api/classify serverless route backed by j-hartmann/emotion-english-distilroberta-base with keyword heuristic fallback. Web Speech API streams interim results so the orb updates mid-sentence.
Also: Supabase for durable session/turn archive, four-level adaptive coaching engine, live Emotion Arc chart (Recharts), pulsing CSS EmotionOrb. Deployed fully serverless to Vercel with no separate backend server.
Rutgers CS + Data Science student, founder of Rolemate, and musician. I build because I like creating things people enjoy, use, and feel helped by. Code is the most accessible way I’ve found to do that at scale; music is the other way I’ve always done it. I play sax, sing, drum, play trumpet, and I’m trying to learn every instrument.
Recent builds: Rolemate, QuantumTrack, PublicWire, and Breathe. I’m interested in AI systems that turn messy human context into useful artifacts: proof maps, audit trails, source-grounded briefs, voice experiences, and evals. I care about products with taste, real usefulness, and serious technical depth underneath.
AI agents, context engineering, eval harnesses, source grounding, provenance, voice UX, speech/prosody interfaces, emotionally adaptive AI, human-in-the-loop workflows, technical recruiting, civic intelligence, AI ROI measurement, startup customer discovery, music-shaped interaction design, and products where code, taste, usefulness, and human behavior meet.
Rolemate: proof-of-skill and interview-prep infrastructure for technical hiring.
QuantumTrack: CFO-facing AI ROI audit system mapping AI spend, vendor claims, infrastructure usage, and operating metrics into finance-verifiable impact.
PublicWire: source-grounded civic intelligence for fragmented local public information.
Breathe: voice-first emotional support concept exploring tone, pacing, hesitation, affect, and conversational safety.