Building RAG Pipelines for Local Llama
2026-01-15Quantized inference, chunking strategies, two-pass retrieval, and streaming chat UX for a fully local documentation assistant.
01
experience
AI Intern
CSG, Government of Karnataka
- Built a multi-agent AI security audit engine deploying specialized persona agents (security, chaos, business logic, compliance) over crawled site graphs to autonomously discover vulnerabilities and generate CVSS-scored reports.
- Engineered accessibility-tree-driven browser automation with 36+ deterministic scanners covering XSS, CSRF, SSRF, JWT, and OWASP Top 10 vectors, plus LLM-powered finding review and deduplication.
- Built an LLM-powered DPR generation platform that runs adaptive multi-turn interviews, extracts structured project data, and generates multi-section PDF reports.
- Developed the end-to-end AI pipeline using FastAPI and self-hosted OpenAI-compatible models, including prompt orchestration, response validation/fallbacks, retrieval-augmented market search, and Whisper-based voice input.
AI/CV Intern
Cyber Physical Systems Lab
- Fine-tuned Vision Transformers (ViT) with LoRA for 5-class diabetic retinopathy grading on MESSIDOR, achieving >80% validation accuracy with stable convergence.
- Implemented parameter-efficient fine-tuning (LoRA), reducing trainable parameters and compute while maintaining competitive performance on limited medical datasets.
- Evaluated model behavior using attention rollout, enabling lesion-focused interpretability and improving failure mode analysis.
AI Intern
FiXitAI
- Built an agentic voice pipeline with LangChain and LangGraph routing for compliance and sales sub-agents.
- Surfaced over 200 high-signal events and reduced decision latency by 10%.
- Engineered sub-second coaching using async streaming and GPT-4 inference routing.
- Productized multi-agent workflows into modular APIs for enterprise use.
Python / ML Intern
AgentStat
- Integrated Whisper-based ML APIs and optimized preprocessing/inference flows.
- Built and profiled computer vision pipelines using PyTorch.
- Applied model-level optimizations to improve throughput and runtime efficiency.
02
featured work
AarogyaAI - Clinical AI Platform
deployedReal-time clinical note generation through a multi-stage STT to LLM reasoning pipeline. Built with PyTorch, Cerebras AI, and Llama 3.1 8B, including multi-agent orchestration and medical entity extraction.
AutoAdapt-Fact System
researchAgentic, self-optimizing fact-checking pipeline with hybrid retrieval (FAISS + TF-IDF), claim extraction with spaCy/BART, and Streamlit-based verification interfaces.
Local RAG Documentation Assistant
activeLocal PyTorch Llama inference with quantized models (GPTQ/4-bit), embedding-based search, streaming chat UI, and containerized deployment architecture for developer documentation.
Diabetic Retinopathy Classification
activeVision Transformer (ViT) with LoRA for 5-class diabetic retinopathy grading on MESSIDOR, achieving >80% validation accuracy. Includes attention rollout for lesion-level interpretability and failure mode analysis.
03
writing
Vision Transformers for Medical Imaging
2025-12-02From CNN baselines to ViT+LoRA: handling class imbalance, parameter-efficient fine-tuning, and interpretability on diabetic retinopathy datasets.
Agentic Fact-Checking with Hybrid Retrieval
2025-11-20Claim extraction with spaCy/BART, FAISS + TF-IDF hybrid retrieval, and self-optimizing verification loops that improve factual reliability.
04
contact
Open to collaboration, consulting, and interesting engineering problems. Reach out directly or use the form.