Akshay Deshpande

Led AI transformation across pharma marketing & drug research — shipping 3 production LLM systems tackling pharmaceutical's hardest challenges: multi-agent orchestration, hybrid RAG & LLMOps at Fortune 500 scale. Ace Performer 2024 & 2025. ▸ Marketing Copilot (Feb 2025–Present) • Architected LangGraph multi-agent system using ReAct + Tree-of-Thought for complex marketing queries • Fine-tuned BERT for pharma NER — drug names, molecules & treatment regimens • Reduced response latency by 40% via Redis caching + dynamic model routing • Deployed on AWS EKS with Terraform-based IaC and CI/CD pipelines → 90%+ reduction in research time — from 2 weeks to under 10 mins ▸ Pharma Analog Discovery (Mar 2024–Feb 2025) • Designed hierarchical agents (planner → retriever → critic → verifier) with persistent memory • Implemented hybrid retrieval — dense embeddings + BM25 + cross-encoder reranking • Applied LLM guardrails for hallucination detection in a regulated pharma environment → 15× acceleration in drug research workflows ▸ MLR Document Intelligence (Jul 2023–Feb 2024) • Built RAG pipelines with token-aware chunking over Medical, Legal & Regulatory documents • Designed Qdrant/FAISS vector search with MMR reranking + metadata filtering • Implemented prompt injection detection & context validation guardrails → 10× faster content review · 40% accuracy improvement

Turned unstructured legal chaos into structured intelligence — NLP and ML pipelines at 22M+ document scale on AWS. Star Performer 2022. ▸ UCC Collateral Extraction (Sep 2022–Jun 2023) • Built PySpark distributed pipelines processing 22M+ UCC lien PDFs • Implemented OCR + document layout analysis converting unstructured PDFs into structured datasets • Designed serverless ingestion via AWS Lambda, S3 & DynamoDB for scalable transformation → Anomaly detection time reduced from days to minutes ▸ Legal Bill Analyzer (Aug 2021–Sep 2022) • Developed XGBoost + FastText models to detect irregular billing entries across invoice categories • Engineered features from 10,000+ historical invoices, tuning precision–recall to minimise false positives • Tracked 15+ model iterations in MLflow for performance monitoring and drift detection → 55% improvement in anomaly detection accuracy · 10× productivity gain for legal ops

Where it started — building deep learning and NLP systems from the ground up, before LLMs made it look easy. ▸ ML Address Standardization (Jan 2019–Aug 2021) • Trained BiLSTM-based NER model to decompose raw addresses into structured fields (street, city, postal code, region) • Deployed model as a containerised REST API using Flask and Docker for production integration • Conducted EDA and data visualisation across large address datasets to identify patterns and edge cases • Built and evaluated multiple model architectures, iterating on precision across diverse address formats → 20% improvement in decision-making efficiency for downstream address-dependent workflows