I am an AI Application Engineer and Data Scientist focused on building performant and reliable LLM applications that drive business efficiency. Specialized in designing complex Agentic AI flows, integrating Real-Time Voice Agents, and deploying scalable Retrieval-Augmented Generation (RAG) pipelines and MLOps observability into modern web environments.
- 🔭 Current Focus: Multi-Agent Orchestration, Voice AI, Strict-CAG / RAG Pipelines, & AI Observability, Privacy-first Edge AI
- 💼 Experience: AI/ML Evaluator Specialist at Mercor (2026), Former Data Scientist Intern at iNeuron.ai, and ML Intern at F(x) Data Labs
- ☁️ Cloud & MLOps: Azure AI Foundry, GCP (Vertex AI), Docker, CI/CD for ML, & Local LLM Deployment
Azure AI Engineer Associate (AI-102)
Skills: NLP, Computer Vision, Generative AI
Azure Data Scientist Associate (DP-100)
Skills: Model Training, MLOps, Data Pipelines
My core toolkit for architecting, evaluating, and deploying production AI systems.
| Generative AI & LLMs | Data Ops & Cloud | Engineering & MLOps |
|---|---|---|
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Focus: Strict Cache-Augmented Generation (CAG), Agentic OCR, Legal Grounding
Built for GenAI Genesis 2026 to protect Ontario renters from bad-faith eviction notices with strict statutory grounding.
- Zero-Hallucination Architecture: Combines Strict-CAG and Agentic OCR to parse and validate Ontario tenancy documents against statutory law.
- Deterministic Validation: Engineered automated document parsing pipelines to flag compliance errors without speculative LLM hallucination.
- Tech: TypeScript, Next.js, FastAPI, Gemini API, Strict-CAG
🔗 https://github.com/DevanshMistry890/tenant-defender
Focus: Real-Time Voice Orchestration, Bidirectional Streaming, Low-Latency VAD
A high-speed conversational voice agent architected for real-time customer support and hospitality workflows.
- Low-Latency Orchestration: Engineered the orchestration layer to handle voice-activity-detection (VAD), interruption handling, TTS/STT synchronization, and tool-calling.
- Bidirectional Audio Streaming: Deployed a custom WebSocket server to stream audio in real time with human-like responsiveness.
- Tech: Python, Gemini AI, ElevenLabs / EdgeTTS, Twilio / WebSockets, ChromaDB
🔗 https://github.com/DevanshMistry890/hotel-voice-agent
Focus: Vector Search, Semantic Retrieval, Context Engineering
A production Retrieval-Augmented Generation (RAG) system that delivers personalized, context-grounded culinary recommendations.
- Vector Search Pipeline: Built an end-to-end RAG architecture using MongoDB Atlas Vector Search and Google Gemini LLM.
- Hallucination Reduction: Implemented semantic retrieval and prompt optimization to maximize contextual accuracy.
- Tech: Python, MongoDB Atlas Vector Search, Gemini API, Next.js
🔗 https://github.com/DevanshMistry890/recipebook
Focus: LLM Observability, Agent Routing Telemetry, MLOps Cost & Risk Tracking
A client-side agent observability and chaos-testing platform designed to monitor multi-step LLM execution flows in real time.
- Execution & Drift Telemetry: Monitors LLM application performance, tracks agent routing decisions, and visualizes real-time execution graphs.
- MLOps Cost & Risk Mitigation: Established foundational tracking capabilities to analyze token cost optimization and failure modes across complex agent workflows.
- Tech: TypeScript, React, WebGPU, MLOps
🔗 https://github.com/DevanshMistry890/Agent-scope
Focus: Multi-Step Agentic Workflows, Local LLMs, Deterministic Tool Calling
An autonomous B2B procurement and contract negotiation system built for privacy-sensitive enterprise workflows.
- Stateful Multi-Agent Negotiation: Orchestrates multi-step procurement and vendor negotiation flows using the Google AI Agent SDK.
- Privacy-First Execution: Integrates open-source local LLMs and deterministic backend tool abstractions within a Next.js microservices architecture.
- Tech: Google AI Agent SDK, Next.js, Local LLMs
🔗 https://github.com/DevanshMistry890/b2b-autonomous-negotiator
