SHAJAM

A deployed platform for physicians to analyze real patient data and conduct myeloma research

Inventor; patent pending through Mount Sinai Innovation Partners.

Open the live platform · Authorized sign-in required

Built for physicians doing research

I conceived, designed and built SHAJAM end to end at Icahn School of Medicine at Mount Sinai. Seven Tisch Cancer Institute oncologists use the deployed platform daily for IRB-approved scientific research. They can define patient cohorts, examine treatment histories and run statistical analyses through natural-language questions.

SHAJAM executes SQL and Python against real patient data, retrieves supporting clinical notes, and produces figures, tables and cohort exports. Its analyses have supported research outputs at the International Myeloma Society and ASH (Blood).

From a research question to an inspectable analysis

  1. Clarify the study. A preflight filter checks the question against a generated map of available data. It runs the question, probes the dataset or asks the physician to resolve a study-design choice.
  2. Build the cohort. Clinical terms are resolved to OMOP concepts and their descendants. The agent reconciles structured records with notes, pathology, radiology and molecular data, counting distinct patients rather than search hits.
  3. Run the analysis. Five MCP servers connect SQL and concept resolution, hybrid clinical-note retrieval, biomedical resources and a Python analysis runtime. The runtime computes statistics on the selected cohort and renders figures in the chat.
  4. Check the result. A separate evaluator is instructed to re-derive quantitative claims. Results retain SQL and source-note provenance so researchers can inspect the evidence and investigate discrepancies.
  5. Use the output. Researchers receive tables, plots and cohort exports for further analysis and manuscript preparation.

Data engineering and scale

The original AIR·MS source warehouse used SAP HANA and OMOP CDM v5.4, spanning 52 tables, 12M+ patients and 14B+ rows. I built a swappable query relay and deployed a curated myeloma dataset on AWS using DuckDB over S3 Parquet, with a connected molecular catalog.

The deployed dataset snapshot covered 11,030 patients, approximately 93M rows and 4.67M clinical notes, indexed into 9.21M passages. It combined clinical records, pathology, radiology, sequencing from four vendors and clonoSEQ MRD data. These are dataset snapshot sizes, not active-user counts or a live feed of the full hospital warehouse.

Agent and retrieval architecture

  • Hermes Agent and Open WebUI: a physician-facing interface with tool use, delegated analysis, clarification and streamed results. Model routing supports Azure and Amazon Bedrock.
  • Five MCP servers: a compact interface to cohort SQL, clinical-note retrieval, statistical analysis, external biomedical knowledge and bioinformatics functions.
  • Hybrid retrieval: Qdrant combines dense and lexical retrieval with rank fusion, Qwen3 embeddings and GPU reranking. Retrieved passages retain their source context.
  • Real statistical computation: pandas, lifelines and matplotlib operate on queried cohort rows in a separate analysis process. Workflows include survival analysis, treatment-episode reconstruction and cohort comparisons.
  • Verification and provenance: concept resolution, explicit data-quality checks, independent re-checking and source-linked outputs make the analysis inspectable.

Deployment and operations

I built the AWS deployment with Terraform, EC2, S3, an Application Load Balancer and WAF, including GPU inference and retrieval services. The delivery workflow uses immutable build bundles, development-to-production promotion, deployment health checks, drift detection and rollback support.

I also implemented column-level PHI filtering, controlled cohort exports, and a 48-event audit taxonomy for tracing the system’s work. The public link opens the sign-in page; patient data and source code remain private.

Research enabled

The platform has supported a 378-patient extramedullary-disease cohort, a 345-patient BCMA CAR-T/bispecific cohort, and immune-recovery analyses for 75 CAR-T patients. It brings cohort definition, evidence review and statistical analysis into one workflow, reducing the repeated handoff between a physician’s question and an analyst’s SQL or notebook.

That research use is central to the project: physicians use SHAJAM to generate analyses from patient data, and those analyses feed scientific work and publications. My publications page lists my research outputs; the platform’s source and patient-level analyses are private.