Hi, I'm Shuban.

Full-time · Bangalore, India · On-site

• Designed a multi-stage LangGraph agentic pipeline (5 agents: Manager, OCR, Reviewer, Enhancer, Review-2) for automated medical document processing on Azure, achieving ~97% extraction accuracy — a 93% improvement over traditional OCR.
• Reduced document processing costs ~80% by replacing Vision LLMs with an optimized agentic architecture; processes 500-page documents in ~1.5 hours at ~$1/document.
• Engineered a cross-service distributed observability framework using OpenTelemetry, Jaeger, Prometheus, and Grafana — implemented end-to-end trace propagation across Azure Service Bus boundaries and ProcessPoolExecutor workers, with custom dashboards tracking LLM token usage per agent, OCR throughput, and queue depth.
• Architected RECON (Real-world Evaluation, Compliance & Observability Network), an internal AI evaluation platform covering RAG, Agent, and LLM evaluation modules with an EU AI Act compliance reporting layer; submitted to CGI FY2026 Innovation Survey.
• Built a black-box RAG validation pipeline for medical APS documents using 5 QGen agents (factual, inferential, adversarial, negation, multi-hop) and domain-specific BioClinical ModernBERT embeddings.

Full-time · Chennai, India · On-site

• Collected a medical gait dataset using a ₹4L Xsens MTw Awinda IMU kit, capturing stair ascent/descent motion from tailbone, knees, and feet.
• Built Python data pipelines to clean IMU signals and extract time- and frequency-domain features (FFT, IQR, mean, median).
• Applied unsupervised clustering (K-Means, Agglomerative, Spectral) for BMI and weight-based classification, with agglomerative clustering performing best and weight-based labels outperforming BMI in several cases.

Full-time · Bangalore, India · On-site

• Owned the FastAPI-based Python backend, building and running production APIs for Scope 3 carbon emissions prediction with real-time and batch inference.
• Designed normalized tables to store emission data and built a vector-based semantic search engine using NLP techniques like NER and word embeddings, hitting ~99% retrieval accuracy on sustainability data.
• Built and deployed stacked ensemble models with TensorFlow & Keras following clean ETL and preprocessing pipelines, consistently achieving 99%+ accuracy and strong F1 scores in production.
• Created advanced Excel carbon accounting tools with heavy edge-case handling — hidden source tables, locked formulas, complex queries, and embedded regression models for precise reporting.