Backend & AI Engineer | LangGraph • FastAPI • Redis | Open source: TensorFlow, pandas, Podman

B.Tech CSE (AI/ML), Lovely Professional University, expected 2027.

I build LLM pipelines, multi-agent systems, and backend services, and I contribute fixes to open-source projects.

• TensorFlow (Google): Fixed float32 underflow in tf.math.sign for small complex inputs using a stable magnitude computation in C++. PR #116986, merged.
• TensorFlow (Google): Replaced exp(x)-1 with expm1(x) in ELU/SELU kernels for better precision on small inputs. PR #116947, approved, pending merge.
• pandas (NumFOCUS): Fixed Categorical.map with a defaultdict, which replaced missing values with NaN instead of using the default factory. PR #65413, merged into the pandas 3.1 milestone.
• Podman (Red Hat): Fixed Quadlet service generator handling of missing keys so volumes are recreated on container restart. Added integration tests and documentation. PR #29375, merged.

Multi-agent startup discovery built with LangGraph, FastAPI, and Redis. Finds founder contacts, drafts short pitches, and sends them through Gmail. Redis-backed rate limiting (10 requests per minute) and a fallback enrichment pipeline.

• Discovery latency reduced from 18s to 4.8s
• Stack: Python, LangGraph, FastAPI, Redis, AsyncIO, Docker, Azure
• Repository

Multi-agent lab report analysis. A 7-node LangGraph pipeline with ChromaDB RAG and a self-correcting CrewAI judge loop.

• MD5 caching reduced response time from 69s to under 1s
• Stack: Python, LangGraph, CrewAI, FastAPI, ChromaDB, Groq, Docker
• Deployed on Hugging Face Spaces with GitHub Actions CI/CD
• Repository

Recommends jobs from resume skills and identifies skill gaps using an LLM for resume parsing.

• Handles 75+ jobs and 60+ skills in the seed data
• Stack: Django, PostgreSQL, Groq API, Docker
• Deployed on Render with GitHub Actions
• Repository

Chrome and Edge extension for real-time resume-job fit scoring on LinkedIn and Indeed, using a Groq LLM and a Cloudflare Workers backend.

• Stack: JavaScript, Groq API, Cloudflare Workers
• Chrome Web Store

XGBoost classifier for phishing URLs, trained on 11,000+ URLs with 58 features.

• 91.38% accuracy, 0.97 ROC-AUC
• Stack: Python, XGBoost, scikit-learn, Streamlit, Docker
• Streamlit app with GitHub Actions CI/CD
• Repository

A Llama 3.2 1B Instruct model fine-tuned with QLoRA using Unsloth Studio on a 2,000-example creative writing dataset, then studied and edited with representation-level behavioral editing.

• Base model: meta-llama/Llama-3.2-1B-Instruct
• Hugging Face model page

Languages: Python, C++, Go, SQL AI and LLM: LangGraph, LangChain, CrewAI, RAG, ChromaDB, Groq API, QLoRA fine-tuning, TensorFlow, scikit-learn Backend: FastAPI, Django, Redis, PostgreSQL, AsyncIO, JWT authentication, REST APIs DevOps: Docker, GitHub Actions, Azure, Render, Hugging Face Spaces

• LinkedIn: linkedin.com/in/sachin-mahajan-cse
• X: @SachinMaha22611
• Email: sachinmahajan903@gmail.com