AI/ML Engineer — Agentic AI, RAG & Voice AI · shipping production LLM systems, not notebooks.

Saakshi — Real-Time Multilingual Voice AI Admission Agent Streaming voice agent on Pipecat + LiveKit (WebRTC/SIP) with OpenRouter function-calling — barge-in interruption, sub-second turn-taking, trilingual (English/Hindi/Gujarati) access via browser and direct PSTN dial-in. Optimized Time-To-First-Audio by streaming sentence-level tokens straight into Sarvam TTS; containerized and auto-scaled on Cloud Run. Pipecat LiveKit OpenRouter Sarvam AI Asyncio Docker Cloud Run

Multimodal Edge AI System Privacy-first, zero-API on-device inference — DistilBERT (96.8% acc.) for text emotion, CNN audio/vision models (~92% each), and a 4-bit QLoRA-tuned LLaMA-3.2-1B, all fitting in 8GB VRAM with no external data egress. Sensor-priority fusion pipeline hits sub-200ms per-modality latency. PyTorch LoRA/QLoRA PEFT 4-bit Quantization OpenCV Gradio

PINNsFormer — Transformer-Based Navier-Stokes PDE Solver Physics-informed Transformer surrogate in DeepXDE solving 2D incompressible Navier-Stokes with mesh-free residual loss. Replaced MLP-PINN backbones with multi-head attention + Wavelet activation to eliminate pressure-field drift across time steps, beating MLP-PINN baselines on temporal consistency. DeepXDE PINNs PyTorch Automatic Differentiation L-BFGS

Explainable Exoplanet Candidate Classification End-to-end ML pipeline over raw NASA photometric data — noise-resilient time-series features isolating sensor bias, ~90% per-sensor accuracy and ~70% cross-satellite generalization via stacking ensemble on a heavily imbalanced dataset. Shipped with a no-code researcher diagnostic dashboard. Scikit-learn Ensemble Learning Time-Series Feature Engineering

AI Chatbot Developer · Admission Committee for Professional Courses, Government of Gujarat — Jan–May 2026 Led end-to-end delivery of a trilingual AI admission assistant — shipped in 30 days, serving 10,000+ applicants statewide. LangGraph multi-agent system over 500+ documents (LlamaIndex RAG + PostgreSQL), AWS Lambda ingestion into Qdrant, streaming text+voice at ~3s end-to-end latency.

AI Automation Intern · Olbuz — Jun–Jul 2025 Architected a multi-agent n8n workflow that replaced a 6-hour, 3-person manual process — 70% reduction in operational overhead.

You're Using LangChain. You Probably Can't Explain How It Works. — Towards AI, May 2026 Architectural teardown of LangChain 1.0 / LangGraph: ReAct loops, StateGraph checkpointing, LangSmith observability, multi-layer agent memory.

Unsloth Just Made Fine-Tuning LLMs a Free-Tier Task — Towards AI, May 2026 Memory-efficient fine-tuning via Unsloth + QLoRA on Qwen3-8B, reproducible on free-tier Colab. 1K+ impressions, 43% read ratio.

• Letter of Appreciation — Commissioner of Technical Education, Gujarat; ACPC chatbot recognized as a model initiative for AI-powered public governance
• NASA Space Apps Challenge 2025 — Global Nominee, Top 1,100 of ~16,000 teams worldwide
• HackTheSpring, GEC Gandhinagar — 🥈 2nd Place (2026), 🥉 3rd Place (2025)
• Published author on Towards AI / Medium — agentic AI & efficient LLM fine-tuning

Information Technology undergrad at L.D. College of Engineering (8.5 CGPA). I've led a 20-member multidisciplinary team on a government-funded AI project, presented solutions to academic, industry, and government stakeholders, and competed in 20+ hackathons — where I've sharpened rapid prototyping, product thinking, and shipping under 24–48 hour deadlines.

I work best where research, engineering, and product intersect, and I'm drawn to problems in AI infrastructure, autonomous agents, and developer tools.

Currently learning: LLMOps — eval harnesses, observability, and cost-aware serving for agentic pipelines.