nidhi bharani

I’m an Applied AI engineer with 9+ years deploying production ML, LLM, and voice AI systems in high-stakes, ambiguous environments. I own the full deployment loop end-to-end—discovery, technical scoping, architecture, build, and production rollout—and I measure success through production adoption, workflow impact, and eval-driven feedback loops.

As a co-founder and AI & Product Lead, I built Emphri, an AI companion platform for people with intellectual disabilities, caregivers, guardians, and disability-support organizations. I designed multimodal experiences (text, voice chat, voice calling, reminders, personalized learning, mood check-ins) and voice-first workflows combining LiveKit peer-to-peer WebRTC calling with OpenAI Realtime API-powered AI voice calls. I also built my orchestration layer with LiteLLM and implemented separate memory, retrieval, and context-management layers using Mem0 and ContextualRAG, along with safety-aware workflows for vulnerable-population deployment.

Previously, as Chief AI Officer & Co-Founder at Bharosa HealthTech, I led AI/product delivery for real-time healthcare voice agents and discovery tools. I architected Mira, a multilingual medical voice agent over PSTN calls using the OpenAI Realtime API, with a low-latency pipeline built on Twilio Media Streams, concurrent async WebSockets, and audio transcoding/resampling—achieving ~500–800 ms end-to-end latency. I enabled natural conversations in 100+ languages with interruption handling and full call lifecycle tracking, built an AI medical scribe with OpenAI transcription and GPT-4o for structured SOAP-style notes across 52+ languages, and delivered document intelligence with GPT-4o vision. I also developed Bharosa Score, an ML-based doctor/clinic ranking system grounded in multi-dimensional trust, outcomes, engagement, record-keeping, environment, and value signals.

Earlier, I contributed as an ML engineer and data scientist—fine-tuning GPT models to 99% accuracy, building evaluation/tracing with W&B/Weave and cost-aware reproducibility, and productionizing clinical decision support like IVF FSH dosage optimization in a Dockerized Django/DRF/AppSmith stack. Outside deployment work, I’ve built privacy-first AI tooling like PromptMask (on-device deterministic PII redaction with Gemma 4) and researched adversarial ASR robustness to understand privacy failure modes—because I want reliable, safe AI that stands up in the real world.