Abhishek Patil

▪ Agentic Pipeline R&D (Current): Developing a multi-agent LLM pipeline targeting 86% → 95% accuracy and 50% → 90% precision — benchmarking Gemini, Claude, and OpenAI across specialised stages: page detection, extraction, consolidation, and reverification.
▪ Self-Service LLM Platform: Architected a self-service LLM orchestration platform that reduced pipeline deployment from weeks of per-team codebase setup to self-service minutes — covering prompt engineering, document ingestion, ground-truth mapping, and automated result assessment across ESG/financial corpora (Climate, Governance, IBOR, PCS, Company Fundamentals). Backend on Google Cloud Run with Cloud Job streaming pipelines for async processing.
▪ Proprietary Orchestration Engine: Bypassed LangChain after empirical scalability evaluations; architected a custom production-grade LLM orchestration engine integrating OpenAI, Anthropic Claude, Google Gemini / Vertex AI. Deploying pipelines via Azure Dataflow and GCP Workflows.
▪ AI Agents: Deployed AI Agents for dynamic keyword generation, leveraging historical seasonal data to improve page detection accuracy and automate metadata tagging across financial document types.
▪ Embedding Model Fine-tuning & Page Detection: Fine-tuned Google text embedding models on domain-specific examples (Vertex AI / GCP) achieving 85% recall and 60% precision for automated document segmentation. Prototyped models using Hugging Face Transformers.
▪ LLM Evaluation Framework: Built comprehensive LLM benchmarking and evaluation frameworks tracking cost, extraction accuracy, and error patterns across datasets — driving model selection and prompt iteration decisions.
▪ Developer Productivity: Use of Codex, Claude Code CLI, Cursor and Copilot IDE during code developments enhancing codebase modularity and team-wide coding productivity across the LLM pipeline workflows.
▪ Technical Leadership: Authored system design blueprints, API specifications, and engineering runbooks as the standard reference across teams; mentored junior engineers on async pipeline debugging and prompt engineering, reducing onboarding time significantly.

• Further developed bilingual AI, user and access customers.
• Led research and technical implementation for ABHA (Ayushman Bharat Digital Health Account)
API integration, ensuring compliance with Indias national health digitisation standards.
• Workflow Automated post-processing workflows by developing executable scripts for email
synchronisation and error-robust file segmentation, reducing manual data-handling errors.

▪ Developed a full-stack application to transform PDF content into image-based visual data, featuring a custom UI for hierarchical document categorisation via Voronoi Tree maps.
▪ Implemented NLP pipelines using TF-IDF vectorisation to extract key semantic terms from research documents for automated word-cloud generation.
▪ Architected the backend infrastructure using Flask and PostgreSQL, solely designing the entity-relationship (ER) models to maintain complex data integrity.
▪ Engineered Directed Acyclic Graphs (DAGs) to efficiently store and navigate relational tree-map metadata, optimising front-end rendering performance.

▪ Developed interactive R Shiny web application to identify and visualise kinship inconsistencies in genomic (GWAS) datasets — application and research called KinVis.
▪ Engineered data pipelines for large-scale GWAS matrix processing with dimensionality reduction techniques to project high-dimensional data onto 2D relational graphs.
▪ Optimised UI performance for real-time visualisation of kin relationships, with the production application successfully hosted and accessible at KinVis.