Arpan Mishra

This is how my professional journey has been until now

Clinical Trial Document Authoring Platform (Fortune 500 Pharma · MS Word Add-in Copilot)

• Problem: Medical writers spent hours synthesising 100+ regulatory documents to produce a first draft — documents reviewed by the FDA where even minor inaccuracies trigger rejection and costly rework. Worked with Medical Writers, Clinical Scientists, PMs, and Solution Architects to build a copilot embedded in MS Word that reduces first-draft time to minutes.
• Built a multi-mode digitization pipeline: fine-tuned Detectron2 (pretrained on PubMed) for client-specific document layout detection; enhanced AWS Bedrock Data Automation + Azure Document Intelligence with custom post-processing for tables, headers, and section formatting; built a section-aware chunking strategy and an LLM skeleton parser (Claude Haiku) that reads every page, summarises each section, and constructs a full document hierarchy for semantically rich indexing.
• Identified that users map source sections to target sections, not text chunks — making conventional RAG the wrong model. Designed agentic section-RAG with a bookmark extraction tool: agent extracts the full document bookmark hierarchy, selects the top 3 relevant sections per query, and returns them as grounded context — significantly improving retrieval precision for long-form regulatory authoring.
• Designed a Planner-Executor multi-agent framework across 12 document types; built a base agent class with LangSmith observability, Portkey for LLM proxy and prompt versioning (managing 100s of prompts and model configs team-wide), and short-term memory. Deployed centralized knowledge as an MCP server on AWS Agent Core; established LLM-as-a-judge eval with golden responses and ACS-framework prompts via LangSmith; human validation checkpoints at critical stages for FDA compliance.

Clinical Query Analysis System

• Problem: Clinical teams had no objective framework to assess query quality — query tagging was entirely subjective. A stratified sample analysis revealed 14% redundancy, significant enough to justify building a systematic classification pipeline.
• Built a query classification framework applying three independent classifiers to every query: Query Type (data mismatch, labelling issue, missing information, etc.), Query Quality (actionable vs. non-actionable — based on clarity, presence of next action, and whether a genuine issue was raised vs. user frustration), and Redundancy Detection against a corpus of ~1.3M historical queries using LLM classification combined with statistical analysis.
• Deployed as an AI validation layer on the live clinical query portal — whenever a user raises a query, the system automatically flags redundancy, assigns a quality score, and tags it with a query type in real time.

Patient Safety Narrative Authoring

• Problem: Original brief: a drafting copilot for patient safety narratives. Identified the real bottleneck was validation — verifying every LLM statement against patient data scattered across disconnected silos. Built an ETL pipeline consolidating adverse event data into a per-patient fact table visualised as a Gantt chart of event timelines — giving medical writers a single source of truth for sign-off and grounding the LLM at generation time.
• Designed a 3-step generation pipeline: (1) Data-to-Facts generation with code-level validation (dates, medication names, adverse events verified programmatically); (2) Narrative Refinement module applying FDA writing guidelines; (3) optional Augmented narrative generation for clinician notes and additional documents. CoT, few-shot, and temporal reasoning prompting; scaled to 600+ patients across 3 trials.

Auto Document Redactor

• Problem: Regulated clinical documents required PII redaction before external sharing — errors risked compliance violations in FDA-governed workflows. Designed a hybrid redaction pipeline: rule-based heuristics for deterministic patterns + fine-tuned spaCy NER trained on clinical PII + chain-of-thought prompting for context-sensitive entities — achieving 98% recall / 90% precision. Built evaluation benchmarks for repeatable compliance testing; enabled confident pipeline iteration without regression.

ISR Entity Detection Pipeline

• Problem: Stakeholders needed to make data-driven ISR sponsorship decisions, but relevant information — drug names, dosages, treatment cycles, endpoints, inclusion/exclusion criteria — was buried in unstructured clinical protocol documents. Built an end-to-end entity extraction pipeline combining spaCy NER, custom fine-tuned entity recognition models, and XGBoost classifiers to structure protocol intelligence at scale for systematic comparison of ISR proposals.

Patient Discontinuation Prediction

• Problem: Client needed early identification of patients at risk of therapy discontinuation to enable timely retention interventions.
• Developed an XGBoost model with features engineered from claims data, patient demographics, and census datasets; used k-fold cross-validation for robust performance estimation and RFE with forward/backward elimination for feature selection; packaged as a reusable cross-client asset enabling client-specific model retraining across ZS engagements.

Recidivism Prediction — Mental Health & Suicide Risk

• Worked with medical data for mental health patients with a history of suicide attempts.
• Modeled the recurrence of suicide attempts using both parametric and non-parametric statistical methods, incorporating medical survey data from VigilanS and identifying factors affecting re-attempt probability.
• Conducted spatial analysis of patient data and incorporated geostatistical spatial autocorrelation into the model to account for regional patterns.

• Worked with satellite imagery and survey data from Census and DHS as part of a global team of 50 change makers.
• Used Landsat 7 & 8 Satellite Images and census data to create a model predicting district-level census variables using a multi-modal, multi-task learning approach.
• Used DHS data and Sentinel images to classify the Asset Wealth Index of clusters across India.
• This project was hosted by World Resources Institute (WRI) and is under UN´s Sustainable Development Goal 8 (Decent Work & Economic Growth).

Gen AI
Multi-Agent Orchestration, RAG, Document Digitization and Parsing, MCP Server Design, LangGraph, LangChain, Claude SDK, LangSmith, Portkey, CoT / Few-Shot / Structured Prompting, LLM Evaluation (LLM-as-a-Judge, ACS Framework), State & Memory Management

Cloud & Infra
AWS (Bedrock, Bedrock Data Automation, SageMaker, Agent Core, EKS, DocumentDB), Google Cloud (Vertex AI, Gemini API), Azure Document Intelligence, Docker

ML
NLP, Layout Detection, spaCy NER, HuggingFace, PyTorch, XGBoost

Languages
Python (expert), SQL, R, Bash

These are some of the personal projects that I have built in the past.

CalorieBot — Nutrition Tracking Agent (2025)

A LangGraph-powered 6-phase nutrition agent deployed as a WhatsApp bot. Users describe meals in natural text or voice and the agent runs item parsing → food search → macro scaling → diary logging → daily summary, end-to-end.

Code

Rossman Sales Prediction

Created a tool to predict the daily sales of any store of the Rossmann drug store chain which is the 2nd largest drug store chain in Germany.

Code

Sentiment Extraction using Bert

Used Bert to detect the sentiment of a given text and further extract the words that best conveys the detected sentiment.

Code

Generating Anime Synopsis using Deep Learning

I used two techniques, LSTMs and then a fine tuned GPT2 for comparing their language modeling capabilities and the results were astounding!

Code

Global Suicide Analysis EDA

I analyzed the global suicide data for 90+ countries from the year 1985 - 2015 in R. Various statistical tests and data visualization techniques were used to explain the data.

Code

Text Analysis Webapp

The purpose of this app is to offer anyone starting off an NLP project a fast and convenient means of exploring the text data cutting down the time between EDA and Modelling.

Code

Rubik’s Cube Rotation Prediction

Predicting the X-Axis Rotation for a given rubik’s cube using Resnet-50. This was part of the AI Blitz Challenge, a hackathon hosted by AI Crowd.

Code

Selfie Filter using CNN

I used a CNN architecture for facial keypoint detection and further used openCV to achieve the desired effect of a sunglass filter which works real time with a webcam.

Code

Here are few of the blogs that I have written related to machine learning, data science and the projects that I have built.

Faster Machine Learning Using Hub by Activeloop

A code walkthrough of using the hub package for satellite imagery

Read

Let’s make some Anime using Deep Learning

Comparing text generation methods: LSTM vs GPT2

Read

Decoding Support Vector Machines

Intuitively understand how Support Vector Machines work

Read

Predicting HR Attrition using Support Vector Machines

Learn to train an SVM model following best practices

Read