AI Research Associate at Vidur Research, working on production RAG systems, LLM fine-tuning, and agentic workflows.
LinkedIn · Portfolio · Medium · The Vector Daily
- Role: AI Research Associate @ Vidur Research
- Education: B.Tech CSE, KIIT University (2023–2027)
- Location: India
- Focus: Production RAG systems, LLM fine-tuning, agentic workflows
- Currently: Building financial intelligence systems with CRAG and RAGAS at scale
AI Research Associate — Vidur Research (Subsidiary of Dreamskrin) Oct 2025 – Present · Remote · India
- Implemented a Corrective RAG (CRAG) pipeline with a retrieval evaluator and knowledge refinement layer, cutting hallucinations 30% across 50+ financial instruments
- Added evidence gating and contradiction detection ahead of generation, reaching 95%+ temporal accuracy with full citation provenance
- Built an autonomous financial research agent performing multi-step retrieval and source-grounded analysis
- Integrated the RAGAS evaluation suite, driving 10–20% retrieval quality gains per iteration across 130+ LLM calls per analysis session
- Reduced inference cost 25% through caching and batching
| Project | Description | Results |
|---|---|---|
| PrepGraph | RAG chatbot with hybrid BM25 + FAISS retrieval, semantic cache, and a query router that selects between Llama 8b and 70b based on query complexity | ↓40% LLM calls · ↓25% latency · ↓30% inference cost |
| Command Copilot | Fine-tuned Phi-2 (2.7B params) with LoRA to translate plain English into Linux commands, running fully offline via Ollama | 85% command accuracy · ↓40% inference time |
| ChatGPT Memory From Scratch | Three-layer conversational memory system (short-term dict, FAISS long-term store, LLM summarization) built without LangChain | No external memory framework used |
| FinVector Research | Embeds financial news with FinBERT and clusters the resulting vector space to detect market regime shifts | Applied on NIFTY 50 · 768-dim vectors · regime transition matrices |
| Reel2Retail | End-to-end CV + NLP pipeline: YOLOv8 detects clothing in video frames, CLIP embeds them, FAISS matches to a product catalog | >75% match confidence threshold |
| The Vector Daily | Automated AI newsletter pipeline — scrapes arXiv, Medium, and HuggingFace, summarizes with an LLM, and ships as HTML daily | 100+ articles/day, live in production |
| SecuFlow | SSH intrusion detection with LLM-based threat analysis, Telegram-based human approval, and automatic UFW blocking | Real-time blocking with explainable threat reports |
- Building an LLM From Scratch: I Trained Word Embeddings on Dostoevsky — Here's What I Found — Towards AI
- Building an LLM From Scratch. Here's Where It All Starts. — Towards AI, May 2026
- My RAG App Was Confidently Wrong — That's When I Found CRAG — Mar 2026