RAG Pipeline
2026Production RAG with a guardrail gate, an evaluator agent, and a reproducible RAGAS harness.
- Python
- FastAPI
- FAISS
- RAGAS
- OpenTelemetry
Satyam · AI/ML Engineer
I turn AI demos into systems people actually use.
I build RAG pipelines, multi-agent systems, and the evaluation harnesses that keep them honest — end to end, from embedding pipeline to Dockerized API. Three internships spent closing the gap between a notebook and a 99% uptime deployment.
- RAG query latencyCloudily Scripts
- 8.2s → 1.7s−79%
- Hallucination rateAsvix · DigiLab
- 18% → 11%−39%
- Deploy timeIPtechhub
- 2h → 15m−87%
- Uptime · 500+ daily queriesAsvix · DigiLab
- 99.2%SLA
Real numbers from shipped work — not vanity metrics. Details in experience.
Selected work
Things I've shipped.
Seven projects — RAG, agents, fine-tuning, ML, and the eval harnesses that keep them honest. Each one has a full case study.
Dual-mode syntax error detection across 5 languages: AST rules first, gradient boosting for the rest.
- Python
- scikit-learn
- AST
- FastAPI
- Streamlit
Four specialized agents turn one product brief into a full campaign — CrewAI execution, LangGraph control flow.
- Python
- CrewAI
- LangGraph
- Groq
- Pydantic
LLM-as-judge that treats position, verbosity, and self-enhancement bias as first-class problems.
- Python
- FastAPI
- SQLAlchemy
- OpenAI
- Anthropic
Parameter-efficient fine-tuning of Llama 3.2 1B for medical Q&A on a single free T4 GPU.
- Python
- PyTorch
- PEFT / LoRA
- Transformers
- HuggingFace
AI Text Detector
2026Explainable AI-text detection on Binoculars cross-perplexity — 100% AUROC on HC3, with a fairness audit.
- Python
- PyTorch
- Transformers
- Streamlit
- Docker
End-to-end MLOps: a 3-model benchmark served via FastAPI with data-drift monitoring.
- XGBoost
- FastAPI
- Docker
- scikit-learn
- Monitoring
Experience
Three internships, one throughline.
Cloud infra → production RAG → hybrid retrieval at scale. Each role pushed a real system closer to reliable.
Jan 2026 — Apr 2026
Remote
AI Developer Intern
Asvix
- Built the embedding pipeline for DigiLab, an AI chatbot on a LangChain + FAISS + Neo4j hybrid RAG stack — 500+ daily queries at 99.2% uptime.
- Cut hallucination rate from 18% to 11% with context-aware response modules and iterative prompt refinement.
- Lifted medical-query relevance 23% using tuned FAISS retrieval plus BM25 hybrid search.
- Wired in LLM evaluation metrics (faithfulness, context recall) to drive prompt iteration with data, not guesses.
- LangChain
- FAISS
- Neo4j
- Hybrid RAG
- LLM Eval
Jun 2025 — Jul 2025
On-site
AI Chatbot Development Intern
Cloudily Scripts
- Built a production RAG pipeline (FAISS IVF128, cross-encoder reranking, BM25 dense retrieval) over 100+ page PDFs at 91% accuracy — cut support tickets 35%.
- Cut query latency 79% (8.2s → 1.7s) with parallel embedding and semantic chunking.
- Dockerized the stack and shrank the image 60% (2.1 GB → 840 MB) for faster deploys.
- Python
- RAG
- FAISS
- Docker
- BM25
May 2024 — Jul 2024
Remote
Cloud Engineering Intern
IPtechhub
- Deployed containerized ML inference on AWS EC2 with auto-scaling — 500+ daily requests at 99.5% uptime, infra cost down 32%.
- Automated CI/CD with GitHub Actions: deploy time down 87.5% (2h → 15m), cold start from 45s to 8s.
- AWS
- Docker
- CI/CD
- GitHub Actions
About
The short version.
I build AI systems that ship. Three internships taught me what the gap between a working demo and a real deployment actually looks like — retrieval that grounds answers, evals that catch regressions, and Docker images small enough to deploy.
At Asvix I built the RAG pipeline behind DigiLab, an educational chatbot handling 500+ daily queries. At Cloudily Scripts I cut query latency from 8.2s to 1.7s on a live PDF RAG system. At IPtechhub I automated deployments from two hours to fifteen minutes.
I also co-authored research on hybrid syntax detection — AST parsing plus a gradient-boosting classifier across five languages, now being prepared for IEEE submission.
My approach is boring on purpose: make it simple, make it work, then make it better.
- Education
- B.Tech, AI & ML — United Institute of Technology (2026)
- Focus
- RAG · multi-agent systems · LLM evaluation
- Research
- Hybrid syntax detection — IEEE submission in prep
- Based
- India · open to remote / global
- Languages
- Python, SQL · English, Hindi
Stack
What I work with.
Tools I've shipped to production — not a wishlist. Depth in the AI/LLM and backend rows.
Languages
- Python
- JavaScript
- TypeScript
- SQL
AI & LLM
- RAG
- Agentic AI
- LangChain
- LangGraph
- CrewAI
- Embeddings
- RAGAS
- LLM Evaluation
- Prompt Engineering
Backend
- FastAPI
- REST APIs
- Node.js
- Microservices
Data & ML
- scikit-learn
- Pandas
- NumPy
- XGBoost
- NLP
Vector & DB
- FAISS
- Pinecone
- PostgreSQL
- MongoDB
- Neo4j
DevOps & Cloud
- Docker
- AWS
- CI/CD
- GitHub Actions
- OpenTelemetry
- Linux
Contact
Let's build something that survives production.
Open to full-time AI/ML roles and freelance work — remote or global. If you're hiring for RAG, agents, or eval-heavy systems, I'd like to hear about it.
Currently open to opportunities — usually reply within a day.