Hi, I'm Balaji Koneti - AI Engineer
AI Engineer · GenAI/RAG · Agentic AI
6+ years shipping production retrieval & evaluation systems on AWS. Measurable gains in relevance, latency, and cost through evaluation-driven ML.
Shipping production AI
with measurable impact.
I'm Balaji Koneti, an AI Engineer (GenAI/Agentic AI) with 6+ years in software engineering. I currently build agentic AI solutions for insurance on Guidewire at Liberty Mutual, after shipping production retrieval and evaluation systems on AWS.
I've delivered measurable gains in retrieval relevance (+22% P@5 on 450 enterprise queries), latency (P95 1.3s to 640ms), and LLM cost (-31%) through evaluation-driven iteration, scalable inference microservices, and reliability guardrails - and cut claims triage turnaround by 55% with agentic workflows orchestrated through n8n.
I hold a Master's in Computer Sciencefrom Northern Arizona University, where I served as a Graduate TA for Machine Learning - leading model explainability reviews with SHAP & LIME across 12 projects and 30+ models.
+0%
Retrieval Relevance (P@5)
0% → +22%
1300ms
P95 End-to-End Latency
1.3s → 640ms
-0%
LLM Cost per Request
Saved 31% per request
0+
Years Engineering
And counting…
The stack behind
production ML systems.
From retrieval pipelines to evaluation frameworks - the tools I use to build, ship, and measure ML at scale.
Programming Languages
GenAI / RAG
LLMs
Machine Learning
Evaluation & Observability
ML Serving & APIs
Cloud & Infrastructure
Data Engineering
Built to solve
real-world problems.
End-to-end ML projects with quantified impact - from NLP-powered assistants to security threat detection pipelines.
Code that ships
in the open.
A year of commits and the repositories behind my work - fetched live from GitHub.
405 contributions in the last year
Academic foundations
that shaped my craft.
M.S. in Computer Science with ML specialization, backed by strong CS fundamentals from India.
Master of Science in Computer Science
Northern Arizona University
- Graduate Teaching Assistant for Machine Learning - led explainability reviews across 12 projects and 30+ models
- Deep focus on ML evaluation, model interpretability, and NLP/LLM systems
- Built reusable SHAP/LIME notebooks adopted as course-wide standard
Bachelor of Technology in Computer Science
Jawaharlal Nehru Technological University
- Strong foundation in data structures, algorithms, and systems programming
- Active in hackathons and technical competitions
- Executive Body Member of Computer Society of India (CSI)
6+ years shipping
production systems.
From production RAG services to fraud detection pipelines - every role measured by real impact.
AI Engineer
Liberty Mutual Insurance
Building agentic AI solutions for insurance on Guidewire - cutting claims triage turnaround by 55%, automating 25+ operations workflows with n8n, and reducing underwriting document review effort by 60%.
- Cut claims triage turnaround by 55% (4.2h to 1.9h) by building agentic AI workflows over Guidewire ClaimCenter FNOL data that auto-classify severity, coverage, and routing using LLM-based extraction with human-in-the-loop checkpoints.
- Automated 25+ recurring insurance operations workflows, saving ~120 analyst hours/month, by orchestrating n8n pipelines that connect Guidewire (PolicyCenter/ClaimCenter) events to LLM agents, notifications, and downstream systems.
- Reduced manual underwriting document review effort by 60% by shipping a multi-agent document-intelligence pipeline (planner / extractor / verifier agents) over policy submissions with confidence-based escalation to underwriters.
- Improved answer faithfulness of the internal insurance knowledge assistant to 96% by building a RAG service over policy and compliance documents with retrieval evaluation gates (RAGAS) enforced before each release.
- Passed internal model-risk review on first submission by implementing agent guardrails - tool allow-lists, PII redaction, action audit logging, and rollback paths - across all production agentic workflows.
Machine Learning Engineer
Nordstrom
Led design and rollout of production RAG services, cutting P95 latency from 1.3s to 640ms, improving retrieval relevance by 22% (P@5), and reducing LLM spend by 31% per request.
- Led design and rollout of production RAG services (LangChain & pgvector), improving retrieval relevance by 22% (Precision@5) on ~450 real enterprise queries via semantic/recursive chunking and hybrid retrieval tuning.
- Cut P95 end-to-end latency from 1.3s to 640ms by separating embedding + generation services, adding Redis caching keyed by (query, tenant, filters), and batching embedding calls.
- Reduced LLM spend by 31% per request by enforcing token budgets, prompt compression, and dynamic routing (retrieval-only / low-risk flows) to smaller models without degrading answer quality.
- Built evaluation & regression pipeline combining Label Studio human review with LLM-as-judge (RAGAS & custom GPT graders) to catch faithfulness/relevance regressions; operationalized as a release gate.
- Productionized inference service (FastAPI & AWS) with health checks, circuit breakers, and explicit abstain/empty-retrieval handling - improved reliability and lowered hallucination in low-confidence scenarios.
Graduate Teaching Assistant - Machine Learning
Northern Arizona University
Led model explainability reviews using SHAP & LIME across 12 projects / 45 students / 30+ models, building reusable interpretability notebooks that reduced time-to-diagnose by 40%.
- Led model explainability reviews using SHAP and LIME across 12 projects / 45 students / 30+ models, identifying leakage and bias patterns and converting findings into concrete feature redesign and retraining recommendations.
- Built reusable interpretability notebooks and visualization templates adopted in course workflows, reducing time-to-diagnose model issues by 40% and improving clarity for non-ML stakeholders.
- Coached teams on evaluation hygiene (proper splits, leakage checks, error analysis) and improved project report quality, reducing rework cycles by 2 iterations.
Software Engineer - Machine Learning Systems
Infosys Ltd
Built fraud-detection pipelines over 1M+ transactional records, improved ETL performance by 43%, and integrated ML outputs into backend services through resilient API patterns.
- Built and operationalized fraud-detection pipelines over 1M+ transactional records, delivering model features and real-time inference via REST API for XGBoost models - enabled faster detection and reduced fraud losses.
- Improved ETL performance 43% through SQL tuning (query refactors, indexing, partitioning) and automated data-quality validation for null spikes and schema drift.
- Integrated ML outputs into backend services through resilient API patterns (timeouts, retries, structured errors), improving reliability and reducing failed inference runs by 15%.
Software Engineer
Nerds and Geeks Pvt Ltd
Owned backend API development, cutting P95 query latency from 450ms to 120ms and automating reporting workflows from 4 hours to 20 minutes.
- Owned backend API development for data-driven applications; improved system reliability by adding input validation, structured logging, and consistent error semantics - reduced production defects by 20%.
- Optimized database schema and SQL queries (indexes, query rewrites), cutting P95 query latency 450ms to 120ms and enabling 35% higher throughput under load.
- Automated reporting/data workflows using Python and SQL, reducing manual reporting from 4 hours to 20 minutes and improving data consistency via repeatable pipelines and sanity checks.
Validated expertise,
industry recognized.
Professional certifications backing hands-on ML engineering experience.
Let's build something
extraordinary together.
I'm always open to exciting collaborations, new roles, and conversations about AI, machine learning, and the future of technology.