I build LLM-powered systems that retrieve, reason, and generate grounded outputs from live data.
Multi-agent orchestration for automated research workflows
Planner → Search → Scrape → Retrieve → Writer → Evaluator
- ~15–20s end-to-end latency across complex multi-hop queries
- 5–8 live sources per query with citation-backed outputs
- 50+ query evaluation using LLM-as-Judge (relevance, faithfulness, completeness)
- Each agent independently testable — production-style modular control
→ GitHub
Grounded Q&A over large document corpora
- Retrieval relevance lifted from 65% to 82% through chunking + embedding strategy
- 10,000+ chunks indexed with FAISS-based semantic retrieval
- Hallucination risk reduced via strict retrieval grounding
- FastAPI inference service — built for deployment, not just notebooks
→ GitHub
AI-powered cold outreach automation
- End-to-end LLM pipeline: job description → portfolio retrieval → personalized email
- ChromaDB semantic search for context selection before generation
- Modular 3-stage architecture (ingestion → retrieval → generation)
- GitHub Actions CI for reliable builds
→ GitHub
ML inference system with production REST API
- P95 latency reduced by 39% (2.7ms → 1.6ms)
- SMOTE-Tomek for class imbalance, scikit-learn + FastAPI
- Containerized with Docker, CI/CD via GitHub Actions
→ GitHub
| Area | Tools |
|---|---|
| LLM Systems | Agentic AI, RAG, LLM Evaluation, Prompting |
| Retrieval | FAISS, ChromaDB, Pinecone, Embeddings |
| Backend | FastAPI, REST APIs, Pydantic |
| Infra | Docker, GitHub Actions, Azure, AWS |
| ML / Data | scikit-learn, Pandas, NumPy |
| Models | GPT-4o, LLaMA 3, Mistral, Gemini |
- More robust agentic systems with stronger evaluation loops
- Production-ready RAG workflows with measurable retrieval quality
- Cloud-deployed AI services on Azure
Portfolio → tvanish002.github.io
LinkedIn → [https://www.linkedin.com/in/anish-tv]
Medium → [ https://medium.com/@anish9]