Applied AI & Data Solutions Engineer
I build governed, role-aware AI systems over SQL and documents — the kind where the hard part isn't getting an LLM to answer, it's making sure it only answers with what it's allowed to know, and can prove its work.
Employees log in to their company's workspace and ask questions in plain English. Answers carry SQL, citations, a chart, an access decision, and a reviewable trace.
Each company gets its own SQL schema and document "brain." Role-based access is enforced at four independent layers — SQL prompt scoping, sqlglot AST table allowlists, ChromaDB metadata filters, and response-level citation checks — so a request can never widen what it's allowed to see.
| Capability | Details |
|---|---|
| Role-scoped access | 4-layer enforcement baked into the agent instance, not a request filter |
| Deterministic analyst layer | 15 business-metric families answered with zero LLM calls |
| Session memory | Multi-turn follow-ups resolve deterministically; re-authorized every turn |
| Admin Health Check Agent | Manually-triggered trace audit that proposes fixes, gated by human verification |
| Full audit loop | Per-employee traces, company-scoped feedback review, trace-leakage auditing |
- 3 synthetic companies, 35-table schema each, on AWS RDS Postgres
- 347+ tests: AST denial, RAG boundary, cross-company memory isolation
- Deployed on AWS ECS Fargate, behind an ALB with ACM-issued HTTPS, frontend on AWS Amplify
- Model gateway across Gemini, Groq, NVIDIA NIM, AWS Bedrock, and local Ollama fallback
Python FastAPI LangGraph Next.js TypeScript PostgreSQL ChromaDB sqlglot AWS (ECS, RDS, Bedrock, ALB) Docker
Processed 534K+ public UCI retail transactions into business-ready analytics across $10.6M analyzed revenue and 4,339 customers.
- Built 5 ML/analytics workflows: 95% F1 churn prediction, K-Means segmentation, Isolation Forest anomaly detection, ARIMA/ETS forecasting — identified 978 at-risk customers
- Accelerated SQL analytics 7.7× by replacing Pandas paths with DuckDB
- Automated executive reports with Groq LLM — cut report generation from 2 hours to 60 seconds
Python SQL DuckDB Scikit-learn Streamlit Plotly Groq LLM
➡️ https://github.com/premsai-pendela/revenueiq-ai
AI/LLM Engineering: Multi-agent orchestration (LangGraph) · RAG (hybrid BM25 + vector, cross-encoder reranking) · role-based access control for LLM systems · deterministic answer routing · Gemini · Groq · NVIDIA NIM · AWS Bedrock
Reliability & Evals: LLM evals, golden tests, retrieval benchmarks, evidence gating, SQL AST validation (sqlglot), trace logging/audit, cost-aware model routing
Languages & Backend: Python · SQL · TypeScript · FastAPI · Next.js/React · REST APIs
Data & Cloud: PostgreSQL (AWS RDS) · SQLAlchemy · ChromaDB · DuckDB · AWS (ECS Fargate, ALB/ACM, RDS, Bedrock, Secrets Manager, CloudWatch, ECR) · AWS Amplify · Docker
Machine Learning: Scikit-learn · Time series forecasting · Customer segmentation · Churn prediction
- Building governed, multi-tenant AI systems where access control is structural, not prompt-based
- Designing deterministic-first analytics so products stay reliable when LLM providers rate-limit
- Deepening AWS cloud deployment experience — ECS, RDS, Bedrock, IAM-scoped infrastructure
📧 nagapremsaip07@gmail.com 💼 https://www.linkedin.com/in/nagapremsai-pendela/ 💻 https://github.com/premsai-pendela
⭐ Open to Applied AI Engineer / AI Data Solutions Engineer roles — feel free to reach out.