I build governed AI systems for regulated industries — banking, healthcare, finance.
My operating doctrine: agents advise, deterministic code decides,
a human acts. Nine years of shipping ML, NLP, RAG and agentic
workflows into production on Azure and AWS.
A governed agentic-workflow pattern from my banking work — shown as a reference architecture, drawn the way it actually runs:
AI Foundry agents write advisories, a deterministic rule makes the call, and every case ends
at exactly one human gate. No agent moves money. Tap a node to see its job.
I'm an AI/ML engineer who has spent nine years taking models out of notebooks and into
production systems that enterprises actually trust — across banking, healthcare,
finance and utilities.
Right now, at Citizens Bank, I design agentic AI for banking operations:
LangGraph orchestration, Azure AI Foundry agents, semantic search over enterprise
knowledge, and the governance layer around all of it — prompt validation, structured
outputs, audit logging, behavior monitoring.
Before that I built RAG and document-intelligence platforms for Cigna,
ML services for financial analytics at Deloitte, and NLP systems for
Duke Energy at Accenture. The common thread: AI that holds up under
regulation, audit and real operational load.
LocationUnited States
CurrentlyAgentic AI Engineer Citizens Bank
FocusAgentic systems · RAG AI governance · MLOps
EducationM.S. Computers & Info. Science Southern Arkansas University
All projects are shown at the pattern level — nothing here includes employer code, data, or confidential information.
Banking · Agentic AIsys.01
Agentic Workflow Automation for High-Stakes Banking Operations
A reference architecture for end-to-end agentic banking workflows: n8n intake, LangGraph orchestration,
Azure AI Foundry advisory agents, a deterministic decision rule and a mandatory human gate —
the pattern drawn in fig.01, proven end to end with monitoring.
LangGraphAzure AI Foundryn8nFastAPIPythonHuman-in-the-Loop
Designed the governance doctrine: agents advise, deterministic code decides, a human acts — no agent touches money.
Built two advisory Foundry agents (transaction review, account status) provisioned once via script, invoked from LangGraph.
Implemented the single-source-of-truth decision rule flagging uncoded transactions and account-status conflicts.
Wired business-readable n8n stages from intake to final document preparation with one merged human-decision node.
Exposed the orchestration as a FastAPI service with a clean 5-dependency footprint, ready for Azure Container Apps.
Banking · RAG · Live in productionsys.02
Enterprise RAG & Semantic Search Platform
Operations teams deal with large volumes of documents, policies and backend
information — much of it searched and validated manually. This platform is the intelligent retrieval
layer: enterprise documents ingested, embedded and served through semantic search and RAG, so LLM
answers are grounded in the bank's own knowledge instead of model memory. Deployed and running in
production.
Azure AI SearchRAGEmbeddingsAzure OpenAILangChainRedis
Built the full pipeline — document ingestion → embeddings → vector retrieval — grounding LLM responses in enterprise knowledge.
Human-in-the-loop controls on sensitive decision points: the model informs, people decide.
Optimized document-intelligence workflows through prompt engineering and retrieval tuning.
Deployed as distributed microservices on Docker, Kubernetes and Azure Container Apps.
Instrumented with Azure Monitor, Application Insights, LangSmith and LangFuse.
Healthcare · RAGsys.03
Healthcare Knowledge & Document Intelligence
Production AI services for a major health insurer — semantic search, intelligent
document processing and predictive analytics, deployed with full MLOps lifecycle management across
regulated healthcare systems.
PythonFastAPIPineconeFAISSMLflowKubernetes
Built enterprise semantic search with vector embeddings, Pinecone, FAISS and Azure AI Search.
Shipped ML models and intelligent search as scalable Python + FastAPI microservices.
Implemented AI governance: model validation, documentation, auditability and responsible-AI deployment.
Automated deployment pipelines with Docker, Kubernetes, MLflow, GitHub Actions and Azure ML.
Improved workflow accuracy via prompt engineering, retrieval optimization and production monitoring.
Desktop tooling · Solo buildsys.04
PDF Editor — Make Any PDF Editable
A desktop application that opens any PDF and makes it directly editable —
click-to-edit text overlay, drag-to-create text boxes, and save-through with the original file as
the source of truth. Built solo, end to end, with an AI-assisted development workflow.
PythonDesktop GUIPDF engineAutomated tests
Click-to-edit overlay on rendered pages plus drag-to-create text boxes anywhere on the document.
Save-through pipeline that writes edits back to a valid PDF and preserves page rotation flags.
Graceful degradation by design: font look-alike fallback for missing glyphs, shrink-to-fit for overlong text.
51 automated checks across the engine and GUI — load, edit, save, dirty-state and folder handling.
Honest limitations documented in the README, including a deliberate no-OCR scope decision.
Utilities · NLP & Predictive MLsys.05
Customer NLP & Predictive Maintenance
Enterprise NLP and ML for a major US utility — complaint classification, intent
detection and equipment-failure prediction feeding customer service and maintenance planning across
operational systems.
TensorFlowScikit-learnspaCyFlaskAWSDocker
Production NLP models for complaint classification, intent detection and text analytics.
Predictive models identifying equipment failures to improve maintenance planning and asset reliability.
REST APIs integrating ML predictions into enterprise customer-service applications.
Hi, I'm Lumen — I answer questions about Tejas's experience, projects, and stack. Try "What did he build at Citizens Bank?" For anything needing Tejas himself, use the contact form below.
Lumen · AI assistant · answers from Tejas's profile
