Healthcare AI systems · Agents · RAG · Medical ASR

Portrait of Promila Ghosh

Promila Ghosh

Senior AI Engineer

Dhaka, Bangladesh

Senior AI Engineer at Golden Harvest Infotech in Dhaka. I design and ship production AI systems: LLM agents, RAG pipelines, and medical speech recognition, with a focus on evaluation, reliability, and scalable architecture.

I architect end-to-end AI for healthcare and enterprise: agents that reason over documents, ASR for clinical speech, and retrieval systems that have to be precise in production. At The Data Dilemma, I work on low-resource AI: datasets and models for Bengali-English clinical speech that generic systems treat as noise. Previously Data Scientist at United We Care, where I built United-MedASR (5.96% WER) and multimodal document agents.

View Projects Download CV Contact Me

Career

Experience

  1. Sept 2025 - Present

    Senior AI Engineer

    Golden Harvest Infotech · Dhaka

    • Led architecture, development, and evaluation of Hppy Claim RCM for USA, UAE, and India markets: 75,000 claims and 82 model runs with 0% cross-tenant leakage.
    • Designed a denial intelligence engine that computes patterns in SQL first, cutting LLM calls from 15 to 1 per analysis with 72-hour tenant-aware caching.
    • Architected a multimodal healthcare document platform: 90% extraction accuracy, 90% agent task success, 85% retrieval Hit@1.
    • Took HealthScan AI to production readiness (2% OCR CER, 40s p95) with a deterministic fallback layer at 100% fallback success.

  2. May 2025 - Present

    Founder & AI Research Scientist

    The Data Dilemma

    • Work on low-resource AI for clinical speech: languages and code-switching that large generic models under-serve.
    • Open-sourced MediBeng, a Bengali-English medical speech dataset with 4K+ samples.
    • Fine-tuned MediBeng Whisper Tiny to 0.98 BLEU (0.3 baseline) and released Medibeng-Orpheus-3b for healthcare translation.
    • Built a healthcare RAG stack (MediRag-Guard) and a streaming LLM interface on FastAPI and Groq.

  3. July 2022 - March 2025

    Data Scientist

    United We Care · Remote

    • Built United-MedASR for medical transcription: 5.96% WER and 7-10x faster inference.
    • Generated 5,000+ hours of synthetic clinical audio with diffusion models and GPT; open-sourced as United-Syn-Med.
    • Shipped a LangGraph multimodal agent for PDFs, scans, and handwriting (~70% accuracy).
    • Built facial and gesture analysis with Redis sliding windows, and a real-time S3 → BigQuery pipeline for an app with 10K+ downloads.

  4. March 2021 - November 2021

    Technical Content Writer

    Qtec Solution Limited

    • Authored 20+ programming courses on machine learning and data structures.
    • Worked with designers and developers to keep material clear and technically accurate.

Education

M.Sc. Engg., Computer Science and Engineering
Khulna University, 2022-2023 · CGPA 3.88/4.00 (Distinction)

B.Sc., Computer Science and Engineering
North Western University, 2017-2021 · CGPA 3.66/4.00

Expertise

Skills

Deep Learning
Neural Networks, Transformers, Large Language Models, Reasoning Models, Vision-Language Models
LLM & ASR
LLaMA, Gemini, GPT, Whisper, Gemma, Retrieval-Augmented Generation (RAG), AI Agents
Tools & Frameworks
Python, Scikit-Learn, PyTorch, LangChain, LangGraph, FastAPI, Kafka, Groq, Fireworks AI, Nebius
Databases & Data
PostgreSQL, MongoDB, BigQuery, Redis, Amazon S3, Qdrant
Cloud Computing
AWS (EC2, S3), GCP (Compute Engine, Vertex AI, Google Cloud Storage)

Selected work

Projects

Hppy Claim

Golden Harvest Infotech

Healthcare RCM has to recover denied claims without leaking data across tenants or inventing a strategy.

I led architecture and evaluation of the Denial Intelligence engine for USA, UAE, and India markets. Patterns are computed in SQL first, then a single LLM call, with tenant-aware caching and a deterministic fallback.

75,000 claims · 82 model runs · 0% cross-tenant leakage · 15 to 1 LLM calls

AI Agents · RAG · SQL · Healthcare RCM

Product

HealthScan AI

Golden Harvest Infotech

Clinical documents arrive as scans, handwriting, and mixed PDFs. Extraction has to hold in production, not only in a demo.

I led architecture through a GO/NO-GO gate: OCR, agent workflows, retrieval, and a deterministic fallback so the system fails closed rather than inventing a strategy.

2% OCR CER · 40s p95 · 90% extraction · 85% Hit@1 · zero safety leaks

OCR · Multimodal models · Agents

Product

United-MedASR

United We Care

Medical transcription is only useful if it is both accurate and fast enough for clinical use.

Transformer ASR for medical speech, with an interfacing path designed for production rather than a research notebook.

5.96% WER · 7-10x faster transcription

Transformers · ASR · Medical speech

Paper

MediBeng Whisper Tiny

The Data Dilemma

Low-resource, code-switched Bengali-English clinical speech is treated as noise by generic models.

I built the MediBeng dataset (4K+ samples) and fine-tuned Whisper Tiny so doctor-patient speech can be transcribed and translated into English for the record.

0.98 BLEU · 0.3 baseline

Whisper · Low-resource ASR · Hugging Face

GitHub Model Paper

MediRag-Guard

The Data Dilemma

Healthcare answers are only as trustworthy as the context they retrieve.

A RAG stack with hierarchical context and ChromaDB, built so retrieval stays attached to source rather than drifting into a fluent guess.

RAG · ChromaDB · Healthcare NLP

GitHub

Medical Document Agent

United We Care

Prescriptions, scans, and handwriting break pipelines that assume a clean PDF.

A LangGraph multimodal agent that reads and classifies medical documents (PDFs, scans, handwritten notes) with a large multimodal model in production.

~70% accuracy on multi-page and handwritten inputs

LangGraph · LMM · OCR

Product

GroqStreamChain

The Data Dilemma

LLM chat is unusable if the model waits to finish before it speaks.

A streaming interface over WebSocket, FastAPI, and Groq so tokens arrive as they are generated.

FastAPI · WebSocket · Groq

Github

Writing & artifacts

Research

Publications

  1. MediBeng Whisper Tiny: A Fine-Tuned Code-Switched Bengali-English Translator for Clinical Applications medRxiv, 2025
  2. Recognition of Sunflower Diseases Using Hybrid Deep Learning and Its Explainability with AI Mathematics, 2023
  3. Multi-labelled Bengali Public Comments Sentiment Analysis with Bidirectional Recurrent Neural Networks Applied Intelligence for Industry 4.0, 2023
  4. Fake News Detection of COVID-19 Using Machine Learning Techniques Springer, 2022
  5. An Empirical Study on Diabetes Mellitus Prediction Using Apriori Algorithm AISC, 2021
  6. Typical and Non-Typical Diabetes Disease Prediction using Random Forest Algorithm ICCCNT, 2020
  7. Human Behavior Analysis using Association Rule Mining Techniques ICCCNT, 2020
  8. A Machine Learning Approach to Identify the Correlation and Association among the Students' Educational Behavior ICCA, 2020
  9. Risk Prediction of Ischemic Heart Disease Using Artificial Neural Network ECCE, 2019
  10. Safeguard: A Prototype of An Application Programming Interface to Save the Disaster Affected People ICCCNT, 2019
  11. A Comprehensive Analysis on Risk Prediction of Acute Coronary Syndrome Using Machine Learning Approaches ICCIT, 2018

Google Scholar profile

Datasets & models

Talks

  • International Conference on Big Data, IoT and Machine Learning, 2021
  • International Conference on Computing, Communication and Networking Technologies, 2019
  • 21st International Conference on Computer and Information Technology, 2018

Get in touch

Senior AI Engineer · AI Architect · AI/ML Lead

Open to conversations about production AI systems. Click the address, or just write.