Sanjana Soma
AI & Software Engineer
Summary
Building intelligent systems with LLMs, RAG pipelines, and full-stack infrastructure.
Education
MS in Computer Science
Experience
Vibho Technologies
- Engineered scalable partner-facing web applications using ReactJS/Redux, building modular UI components that improved load performance by 25% and reduced development redundancy by 20%
- Diagnosed and resolved 30+ production integration and performance issues monthly, improving reliability of distributed client-server interactions
Mindgraph Technologies
- Developed scalable data ingestion modules for a Customer Data Platform processing 90M+ records, enabling large-scale analytics for enterprise clients
- Optimized SQL queries and backend data pipelines, reducing data retrieval latency by 18%
Projects
AI Cold Email Tool
- Built a full-stack outreach tool using FastAPI and Claude API for personalized email generation across 4 tones (Formal, Conversational, Story-Driven, Data-Driven), with Groq LLaMA 3.3 70B as a resilient fallback for provider failure
- Integrated DuckDuckGo for real-time prospect research and Notion API for outreach tracking
RAG pipeline using Groq LLaMA 3.3 70B and HuggingFace embeddings to produce audience-specific summaries from PDF documents across 4 modes, with structured citation grounding to prevent hallucination
- Implemented Zod schema validation with auto-repair to self-correct malformed LLM JSON outputs, ensuring pipeline reliability in production; deployed live on Vercel
Arizona State University
- Engineered a custom multi-agent data synthesis pipeline using LLaMA-3.1-8B-Instruct on the SOL Supercomputer, generating 10,000-word reasoning benchmarks through iterative prompting and a QA modifier agent that stripped answer-leaking context from questions
- Built and manually verified a dataset of 15 long-context narratives and 108 QA pairs covering arithmetic, temporal, and logical reasoning; ran comparative accuracy audit across GPT-4, Perplexity AI, and LLaMA
- Identified a “Lost in the Middle” retrieval bottleneck: model accuracy dropped significantly when key information was positioned near word 2,500 in 10,000-word contexts, consistent across models
Technical Skills
Skills: AI / ML, RAG Pipelines, LLM Evaluation, Prompt Engineering, NLP, HuggingFace, Groq, Frameworks, FastAPI, ReactJS, NodeJS, Django, TensorFlow, Pandas, PySpark, Programming, Python, JavaScript, SQL, C++, Java, C, Systems & Tools, Git, REST APIs, Distributed Data Pipelines, Vercel, Zod, Linux/Unix