Yashwanth
Challa
I build production AI for regulated businesses — putting models where they read and reason, and deterministic engineering where the business needs certainty and an audit trail.
Engineering judgment, not just model output.
Applied ML / GenAI
Agentic systems
Data engineering
Domains
Insurance & benefits
Financial data
Document AI
I'm an AI & Data Engineer whose work has moved steadily from large-scale data infrastructure into applied ML and production GenAI. I've built multi-agent orchestration systems, a Text-to-SQL engine with automated evaluation guardrails, and AI support agents that handle thousands of real inquiries a month — systems people depend on, not demos.
That data-engineering foundation is what makes the AI work trustworthy. I treat model output like any other data source: something to validate, monitor, and cost-control, never blindly trust. My guiding principle is simple — use AI where it's genuinely better, and deterministic code where a regulated business needs certainty.
I care most about the intersection of domain and engineering: insurance and benefits reward people who understand both the business rules and the systems that enforce them. That's the work I want to do — AI that measurably moves a business metric, built to be audited.
AI-Powered Commercial Auto Underwriting Pipeline
The full interactive system architecture — high-level design, the six-step pipeline, the deterministic rules engine, data flow, infrastructure, and security.
Production systems, from streaming pipelines to agents.
Owning the data and AI layer of a venture-backed insurance-tech SaaS platform, from IRS tax-compliance logic to multi-agent AI.
- Designed and implemented the end-to-end data integration architecture for a $500K venture-backed SaaS platform on AWS, translating complex IRS tax-compliance requirements (ICHRA) into scalable data workflows — increasing administrative efficiency by 40% and reaching 100,000 insurance agents within six months.
- Engineered a secure Text-to-SQL engine with multi-layer query-validation logic to autonomously generate complex financial reports from production databases, achieving 99% data integrity via an automated LangFuse evaluation framework with semantic guardrails.
- Architected a multi-agent AI system with LangChain, LangGraph and LangGraph MCPMultiServerClient — connecting Claude/Copilot agentic harnesses with MCP Servers and specialized sub-agents, plus custom routing over Qdrant vector search and Gemini (Google ADK) — reducing processing costs and improving response reliability.
- Built Python Eligibility Engines and Affordability Calculators performing automated validation across 11 employee-classification rules, enforcing ACA compliance and eliminating manual review overhead.
- Developed ChatGPT-based support agents handling 5,000+ monthly policy inquiries against backend data systems, cutting operational overhead by 30%.
- Designed and optimized MongoDB schemas for high-traffic Open Enrollment across 50,000+ lead records; automated policy tracking and reconciliation via Zoho CRM customization.
Applied NLP and deep learning to unstructured financial text to produce signals used in asset-management decisions.
- Built and validated pipelines to extract, transform and analyze large volumes of unstructured financial text (earnings-call transcripts, Reddit, Twitter) using NLP and deep learning — producing ethics risk scores used in asset-management decisions.
- Developed custom Word2Vec models and supply-chain risk lexicons via a proprietary CBOW approach to map ethical-risk signals across NASDAQ-listed firms.
- Trained CNN classifiers (TensorFlow, PyTorch) on social-media datasets to 85% accuracy in signal detection; managed end-to-end ingestion via PRAW and Pushshift APIs.
Built real-time, high-volume integration pipelines and the infrastructure to keep them reliable and secure.
- Designed real-time integration pipelines ingesting 10 million raw records from 12+ heterogeneous sources into Hadoop using Kafka, Elasticsearch and custom ETL — handling semi-structured, schema-variable data at scale.
- Built Apache NiFi pipelines processing 400GB of data daily with configurable Lookup services for automated validation, deduplication and integrity enforcement.
- Authored Python integration scripts and a HashiCorp Vault API library to enforce enterprise security across pipelines and automated workflows.
- Optimized SQL report queries in QlikSense dashboards, cutting report generation time by 85%; orchestrated a Kubernetes cluster from scratch (Minikube) scaled to diverse production workloads.
- Engineered a PDL Management Tool in .NET Core 3.0 with Selenium end-to-end automation, streamlining email-administration workflows for 500+ employees.
Computer-vision research on embedded hardware, published at IEEE conferences.
- Developed a driver drowsiness detection system using Raspberry Pi, OpenCV and Dlib; published findings in two IEEE conference papers (TENCON 2019, TENSYMP 2020), garnering 7+ citations.
Open-source systems, built to production standards.
The Knight underwriting pipeline is featured above. These are two more systems I've built end to end.
VegaRAG
Open-source RAG agent builder deployed on AWS Fargate. LangGraph agents with intent routing (RAG / Text-to-SQL / casual), MCP Servers via LangGraph MCPMultiServerClient, Pinecone with namespace-per-tenant isolation, DuckDB in-memory analytics, Amazon Bedrock Nova inference, LangFuse tracing, and DynamoDB single-table design — all behind one ALB across three independent ECS microservices, with real-time SSE token streaming and a zero-trust IAM task-role model (zero hardcoded credentials).
Video Summarization & Synthesis
An AI-driven content pipeline combining SSD for face detection, Wav2Lip for lip-sync, and TF-IDF / Gensim / NLTK for text summarization — served through a Streamlit interface. A study in stitching multiple specialized models into a single coherent multimodal workflow.