Skip to content
View purvanshh's full-sized avatar

Highlights

  • Pro

Block or report purvanshh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
purvanshh/README.md

Purvansh Sahu

Backend · AI/ML · Machine Learning


About Me

  • 🎓 4th-year Undergraduate CS Student at Scaler School of Technology & BITS Pilani (2023 — 2027)
  • 🔬 Machine Learning Research Intern at IIT Madras (Jan 2026 — Present)
  • 🛠️ I build production-grade systems combining backend engineering with ML research, multi-agent orchestration, and observability
  • 📍 Based in Bangalore, India
  • 🌐 Portfolio: purvanshsahu.vercel.app

Experience

Machine Learning & Python Intern | Indian Institute of Technology Madras (Jan 2026 — Present, Remote)

  • Designed config-driven evaluation pipelines and conducted ablation studies on CNN-based aerial plant disease detection using drone imagery under distribution shift, collaborating with ICAR -- National Research Centre for Banana on field constraints.

Featured Projects

Webhook-driven multi-agent system orchestrating Style, Logic, and Security review agents in parallel to post confidence-scored inline feedback on GitHub PRs.

  • Tech: Python, FastAPI, Celery, Redis, ChromaDB, tree-sitter, Docker, GitHub Actions, Prometheus

Incident reasoning system using LangGraph for multi-node state orchestration, Qdrant RAG dense retrieval, auto-degrading local model fallbacks, and token-gated tool execution safety.

  • Tech: Python, LangGraph, FastAPI, Qdrant, Celery, Redis, Ollama, PostgreSQL, Docker, Prometheus

Full-stack AI gateway with cost-aware routing between local OSS (Qwen2.5) and frontier models, sub-300ms content safety via Llama Guard 3, PII redaction, and a GPT-4-as-judge evaluation harness.

  • Tech: Python, FastAPI, Next.js, Qwen2.5-0.5B, GPT-4, Llama Guard 3, Langfuse, Prometheus, Docker

Credit decision engine leveraging XGBoost with isotonic calibration (AUC-ROC 0.976) and SHAP explainability, featuring KS-based distribution shift detection and transactional audit trails.

  • Tech: Python, XGBoost, FastAPI, Celery, Redis, PostgreSQL, Docker, SHAP, Prometheus, GitHub Actions

Layout-aware multimodal document intelligence system extracting structured JSON from unstructured PDFs using LayoutLMv3 and PaddleOCR, achieving 0.581 F1 score.

  • Tech: Python, PyTorch, LayoutLMv3, PaddleOCR, FastAPI, Docker, Pydantic, NetworkX

Technical Skills

Category Skills & Technologies
Languages Python, SQL, TypeScript, JavaScript, HTML5, CSS3, Bash
Frontend React, Next.js, Tailwind CSS, Shadcn/ui, Material UI, Redux Toolkit, TanStack Query (React Query), React Hook Form, Zod, Framer Motion, Axios
Backend & Infrastructure FastAPI, Celery, Redis, PostgreSQL, MongoDB, Docker, Kubernetes, GitHub Actions, CI/CD
ML & AI Systems LangGraph, RAG Pipelines, Multi-Agent Orchestration, PyTorch, Transformers, YOLOv8, XGBoost
Observability & Tools Langfuse, Prometheus, Grafana, MLflow, SHAP, tree-sitter, Ollama, Git

Connect With Me

Pinned Loading

  1. PRGuard-AI PRGuard-AI Public

    Multi-agent GitHub PR review system with parallel Celery workers (Style, Logic, Security agents), AST-based code parsing via tree-sitter, confidence arbitration with IoU-based span clustering, and …

    Python

  2. Github-Rag Github-Rag Public

    RAG pipeline over GitHub repositories, indexes codebases into a vector store and enables natural language querying over code, PRs, and documentation.

    Python

  3. AuditLend AuditLend Public

    Production-grade loan eligibility and credit decisioning backend featuring a configurable rule engine, borrower risk scoring, audit trails, and transactional outbox pattern for reliable event delivery

    Python

  4. Document-Intelligence-Engine Document-Intelligence-Engine Public

    Document understanding pipeline using LayoutLMv3 and PaddleOCR for structured data extraction from complex PDFs and scanned documents. FastAPI backend with Docker deployment.

    Python

  5. Transformer-from-scratch Transformer-from-scratch Public

    A from-scratch implementation of the Transformer architecture to understand attention mechanisms, embeddings, and training dynamics.

    Jupyter Notebook

  6. Portfolio-Machine-Learning Portfolio-Machine-Learning Public

    A curated collection of machine learning projects covering data preprocessing, model training, evaluation, and experimentation workflows.

    JavaScript