Building systems that retrieve, reason, and ship to production.
I'm Mihir Rathod — an AI/ML engineer who builds RAG pipelines, fraud detection models, and NLP-powered automation that actually reaches a deployed URL, not just a notebook.
From notebook to deployed endpoint.
I'm a final-year B.Tech ICT student at PDEU, Gandhinagar, but most of what I know didn't come from a syllabus — it came from shipping things that had to actually work on real data, for real users, under real constraints.
At Brainybeams, that meant pushing a fraud classifier past 90% accuracy on a dataset where fraud was 0.17% of transactions — the kind of class imbalance that breaks naive models. At Bytesbizz, it meant building a 6-module hiring pipeline that had to parse messy, inconsistent resume PDFs and still produce a reliable ATS score.
What ties my projects together is a refusal to stop at "the model works in Colab." I care about the retrieval layer, the inference latency, the deployment target — the parts that decide whether an ML idea becomes a product or stays a demo.
Plotted, not just listed.
Every skill below is positioned by how I actually use it — clustered near the techniques and tools it pairs with on real projects. Click a point to inspect it, the way you'd inspect a vector in an embedding space.
to see usage details,
projects, and confidence.
Two internships, two different failure modes solved.
Each one forced a different kind of rigor — language ambiguity in hiring data, class imbalance in fraud data.
- Built a 6-module AI Hiring Copilot covering resume parsing, ATS scoring, gap analysis, question generation, answer evaluation, and hiring report generation.
- Implemented an NLP resume parser with spaCy that extracts 20+ skills from PDFs, plus an ATS scoring engine comparing resumes against JD requirements using set-intersection logic.
- Generated personalized interview questions via Groq LLaMA and evaluated candidate answers across technical accuracy, approach, and communication.
- Deployed a 4-stage recruiter dashboard with Hire / Maybe / Reject logic, using Streamlit session-state for a stateful multi-step workflow.
- Developed a fraud transaction classifier with Scikit-learn and Random Forest on 284,807 transactions (0.17% fraud rate), achieving F1 0.84, ROC-AUC 0.918, precision 85%, recall 84%.
- Performed end-to-end feature engineering, outlier handling, and class-balancing via SMOTE to improve minority-class fraud sensitivity.
- Benchmarked 4 models — Logistic Regression, Decision Tree, Random Forest, XGBoost — on precision, recall, F1, ROC-AUC, and confusion matrix to select the production-ready model.
- Deployed a real-time fraud detection Streamlit app with a direct inference pipeline and HIGH/LOW risk scoring at a 70% probability threshold.
Four systems, each shipped end-to-end.
Click any project to open the full breakdown — architecture, challenges, and what I'd build next.
Formal foundation, informally exceeded.
B.Tech, Information & Communication Technology
Diploma, Computer Engineering
Let's build something that ships.
Open to AI/ML Engineer, Data Scientist, and GenAI roles — full-time or freelance.