Kiran Babu Athina

Texas A&M AISLS Lab · College Station, TX

• Achieved 98.7% Recall@5 across 3 domain-specific knowledge graphs with an agentic RAG system on TAMU HPC (2× NVIDIA RTX 6000), using a query-routing layer over 3 fine-tuned KG retrievers.
• Cut general-purpose cloud API calls to zero for precision nutrition queries by fine-tuning Gemma 4 12B with LoRA on food-systems corpora aligned with SDG 2.
• Reached 95% Hit@5 and 0.925 MRR on a 1,000-question expert-curated eval set with a FastAPI RAG backend on Azure App Service (hybrid BM25 + FAISS over 120,000+ chunks).
• Built a chunk-level scoring and citation pipeline inside the inference server: 86% source precision, 95% source recall.
• Eliminated cold-start 503s, with zero failed session initializations, through a React chat frontend with race-safe bootstrapping, streaming UX, and path-triggered CI/CD on Azure.

Aggie Research Program · Texas A&M University

• Led teams of 3 undergraduate researchers per project, owning end-to-end execution aligned with faculty research goals.
• Scoped projects, broke research objectives into weekly tasks, and mentored students in applied ML, data analysis, and research practice.
• Served as primary point of accountability, reporting progress, risks, and results to the supervising professor weekly.

Texas A&M Department of Animal Science · College Station, TX

• Eliminated Azure OpenAI calls for livestock terminology queries by fine-tuning LLaMA 3 with LoRA/PEFT on a proprietary animal science corpus.
• Delivered grounded, citation-backed answers with a RAG pipeline on Azure AI Search + Azure OpenAI, validated by faculty review across 3 semesters.
• Built a sensor-driven digital twin (MQTT + Mesa agent-based model) for minute-level per-animal energy tracking and monthly methane forecasting.

Tata Consultancy Services · Hyderabad, India

• Cut loan processing time by 30% with classification models on Azure ML using AutoML selection and statistical feature engineering.
• Eliminated 30% of manual hardware log triage and sped up issue resolution by 18% with a TF-IDF + Random Forest NLP pipeline over 500,000+ daily telecom logs.
• Reached 92% pass/fail accuracy on hardware label QC across 10,000+ devices/month with a multimodal ResNet-18 + BERT verification system.
• Processed 100M+ sensor records/day for fault-prediction models with a PySpark ETL framework on Hadoop.
• Reduced post-deployment performance decay by 22% with drift dashboards and retraining triggers, and cut deployment turnaround by 30% with Docker + Kubernetes auto-scaling.

Tata Consultancy Services · Hyderabad, India

• Achieved 98.5% fraud classification accuracy on KYC anomaly detection with a multimodal PyTorch model fusing 1D CNN audio and 2D CNN document-image embeddings.
• Reduced onboarding and handoff time by 25% by writing end-to-end documentation of model architectures, assumptions, and evaluation results.