NCR Atleos — Associate Software Engineer
Oct 2025 — Present- Building internal AI/automation tooling on production LLM and retrieval services.
AI Backend Engineer
LLM Systems · RAG Pipelines · Agentic Workflows
I build production LLM systems — retrieval that stays fast and cheap when it can, and goes agentic only when a question genuinely needs it.
Submit a query — the gate takes the cheap deterministic path when it can, and spends on the agent loop only when a question needs it.
AI Backend Engineer specializing in production LLM systems — RAG pipelines, hybrid retrieval, and agentic orchestration.
About 17 months building and shipping AI systems: an enterprise AI analytics platform (async ingestion, hybrid retrieval, an agentic query engine) and a live ML product with 56 real users on the Chrome Web Store. Strong in Python, FastAPI, LangGraph, Azure OpenAI, and vector/graph retrieval — and opinionated about doing the expensive thinking at ingestion so queries stay fast.
FastAPI · LangGraph · OpenSearch · Azure OpenAI · Celery/Redis · PostgreSQL · DataFusion · Docker
A 13-stage async ingestion pipeline (Celery + Redis) that turns messy ERP/CSV into hybrid OpenSearch indexes (BM25 + HNSW) — intelligence computed at ingestion, never at query time. The query engine reranks with RRF and gates on a deterministic planner: high-confidence queries take a fast DataFusion path (~1–2s, zero LLM calls); multi-hop queries fall back to a LangGraph agent loop bounded to ≤14 tool calls. Value-validated joins via GIN-indexed array overlap, multi-tenant isolation at every layer.
TODO: replace with corrected freelance copy
Next.js · FastAPI · LangChain · Neo4j · PostgreSQL · Azure OpenAI · AWS S3
Agentic document intelligence. Extracts entities into a Neo4j knowledge graph for context-aware Q&A across a whole workspace — catching cross-document connections plain vector search misses. A retrieval-grading loop runs 3–4 queries per question over 6–7 iterations, scoring by confidence to auto-select the best answer.
Python · Flask · React · AWS EC2 · Docker
Real-time phishing detection. An ML model at 95%+ accuracy served behind a live REST API (EC2 + Docker), shipped as a Chrome extension that flags phishing while you browse — 56 real users on the Chrome Web Store.
Next.js · Python · LangChain · Gmail API · Vercel
An AI-first personal finance tool. Reads credit-card data from Gmail and uses an LLM tool-selection engine that autonomously picks the right financial tool per query — no fixed query structure. Roadmap: real-time stock analysis and AI-generated portfolio insights.
Open to roles and interesting problems in AI backend and LLM systems. Fastest way to reach me:
asb.bharath601@gmail.com