AI/ML Engineer & Researcher — from vision robustness evaluation to production LLM systems 🚀

---

🧠 About Me

class SarthakChauhan:
    def __init__(self):
        self.role = "AI/ML Engineer & Researcher"
        self.education = "B.Tech CSE (AI/ML) @ Bennett University"
        self.achievements = "CGPA: 9.42/10.0 | Dean's List (Top 10%)"
        self.location = "India 🇮🇳"
        
    def current_work(self):
        return [
            "🔬 Vision model robustness: benchmarking 12 architectures across IN-Val/V2/R/A (ECE, NLL, per-class dispersion)",
            "🚗 Fog-highway dehazing benchmark: 10 architectures, 15–20 dB PSNR gap finding (DICCT 2026)",
            "🏫 Production RAG pipeline @ Cograd: 50+ teachers, 6 schools, 42% prep-time reduction",
            "💬 Hinglish abuse detection: XLM-R + BiGRU, F1 0.866 on 700K posts (IEEE AICAPS 2026)"
        ]
    
    def skills(self):
        return {
            "AI/ML": ["Deep Learning", "NLP", "Computer Vision", "RAG", "PINNs"],
            "LLM Stack": ["LangChain", "LlamaIndex", "CrewAI", "AutoGen", "LangGraph"],
            "Frameworks": ["PyTorch", "TensorFlow", "Hugging Face", "FastAPI"],
            "MLOps": ["Docker", "MLflow", "W&B", "ONNX", "TensorRT"]
        }
    
    def fun_fact(self):
        return "I think my GPU works harder than I do 😄"

---

🔬 Research Focus

Distribution Shift & Model Calibration Investigating how natural and rendition-based shifts expose calibration failures in vision models. Found training recipe dominates over architecture family: ResNet-50-V1 (ECE=0.039) vs V2 (ECE=0.410) at comparable Top-1 accuracy.

Evaluation Beyond Average Accuracy Building benchmarking frameworks that measure worst-group robustness, per-class dispersion, ECE, and NLL across architecture families (ResNets, ViTs, Swin-T, ConvNeXt, MaxViT) on IN-Val, IN-V2, IN-R, IN-A.

---

🚀 Featured Projects

DataWhiz Text-to-SQL System with Multi-Agent Orchestration 🗃️ Handles 200+ table databases with GPT-4o + LangChain 🎯 35% error reduction via vector schema retrieval over full-schema prompting 📊 3.2× faster insights with LIDA auto-visualization (N=12 user study) ☁️ Deployed on Azure with CI/CD pipeline FastAPI LangChain DuckDB Neo4j Azure	StreamMind Real-time AI Doubt Clustering for Live Classes ⚡ 6-stage async pipeline with dedicated Redis workers per stage 📉 68% reduction in instructor response time (200-doubt benchmark) 🔍 pgvector ANN search collapses semantic duplicates before answer generation 🔴 WebSocket layer supports 100+ concurrent doubts on YouTube Live FastAPI Redis pgvector WebSocket LLMs
Medha AI Enterprise RAG System @ Cograd (Team Project) 🏫 Deployed across 50+ teachers in 6 schools ✅ 78% of content required minimal editing ⚡ 2.5–3.5× latency cut via asyncio parallelization + SSE stream merging (~1s TTFT) 💰 25–30% LLM cost reduction via prompt compression & quantization Qdrant MongoDB FastAPI PostgreSQL Redis	Aurigen AI Jewelry Design Studio 💎 Fine-tuned SDXL via LoRA (FP16, 10K steps) on self-curated 6,157-image dataset 🎨 ControlNet Canny preserves geometric constraints where vanilla SDXL drifted ⚡ 3.9× latency reduction (8.2s → 2.1s) via attention caching + FP16 SDXL ControlNet LoRA PyTorch Streamlit

_{📌 Ongoing Research: Vision Model Robustness Evaluation — benchmarking 12 architectures (ResNets, ViTs, Swin-T, ConvNeXt, MaxViT) across IN-Val, IN-V2, IN-R, IN-A · Measuring ECE, NLL & per-class dispersion · W&B Report ↗ · PyTorch · In Progress}

---

🎯 Skills

🧠 AI/ML & Research Machine Learning, Deep Learning Natural Language Processing (NLP) Computer Vision (YOLOv8, Dehazing, Detection) Transformers, LLMs, RAG Systems Physics-Informed Neural Networks (PINNs) Diffusion Models (SDXL, ControlNet) Optimization, Feature Engineering, Statistical Modeling	🤖 LLM & Agents LangChain, LlamaIndex, LangGraph AutoGen, CrewAI, JinaAI Prompt Engineering & Retrieval Optimization Multi-Agent Systems for SQL, Automation & Pipelines Vector Search & Embeddings OpenAI API Integration
📚 Frameworks & Libraries PyTorch, TensorFlow, Hugging Face HuggingFace Diffusers, SciPy scikit-learn, OpenCV, NumPy, Pandas FastAPI, Streamlit, DuckDB, MongoDB Qdrant, Neo4j, PostgreSQL, MySQL	⚙️ MLOps & Systems Docker, MLflow, Weights & Biases ONNX, TensorRT (FP16/INT8 optimization) GGUF Quantization, Model Compression Azure, GCP, Linux Grafana, Prometheus, CI/CD Experiment Tracking, Profiling & Deployment CUDA, LaTeX

---

🛠️ Tech Stack

💻 Languages

🧠 Deep Learning & AI

🤖 LLM Ecosystem

🚀 Backend & Deployment

📊 MLOps & Optimization

🔧 Tools

---

📊 GitHub Profile Stats

🔥 Streak Stats

📈 Contribution Graph

---

🐍 Contribution Graph

---

🏆 Achievements

🥇 Hackathons & Competitions	🎓 Academic	📜 Certifications
Amazon ML Challenge 2024 Top 0.5% (409/74,823)	Dean's List Award Top 10%	IBM Machine Learning
IIT Bombay Convolve Top 50/4,189 Teams	CGPA: 9.42/10.0	Deep Learning Specialization (Andrew Ng)
Kharagpur Data Science Semi-finalist	Published @ IC3SE 2025	GenAI with LLMs
		AI Agents Intensive — Google × Kaggle 2025

---

📚 Research & Publications

📄 "Hinglish Abusive Comment Detection Using Transformer-Based Models" (First Author) Accepted at AICAPS 2026, IEEE Kerala Section — XLM-R + BiGRU, F1 0.866 on 700K+ code-mixed posts

📄 "Image and Video Dehazing for Dense-Fog Indian Highway Scenarios" (First Author) Accepted at DICCT 2026 — Benchmarked 10 dehazing methods; identified 15–20 dB PSNR gap between synthetic benchmarks and real dense-fog conditions

📄 "Deep Learning-Based Brain Tumour Identification" (Second Author) Accepted & Presented at IC3SE 2025, IEEE UP Section — Residual CNN, 97.10% accuracy at 5M parameters

---

💭 Dev Quote

---

🤝 Let's Connect!

I'm always excited to collaborate on innovative AI/ML projects!

💼 Open to: Research Collaborations | Open Source | AI/ML Internships

📧 Reach me at: sarthak4156@gmail.com

---