Aamer Jalan

AI-native EHR for private pay and out-of-network behavioral health clinicians. In production, with compliance enforced in code instead of in policy documents.

• Built the AI documentation agent on Claude, review-first, with clinical safety guardrails
• Turned HIPAA and 42 CFR Part 2 into enforced code paths: ABAC over FHIR R4, hash-chained audit
• Shipped the billing engine: Stripe autopay with exactly-once charges, 837P claims, ICD-10 validation

Open-source clinical backend for AI health apps. Canonical FHIR R4, cited search, and governed agent proposals, so an agent reads scoped evidence instead of a raw-FHIR dump.

• Put governance in the database: row-level security that fails closed from the first row
• Propose, approve, commit, so an agent can draft a clinical write but never approve its own
• Passes 133/133 HL7 SQL-on-FHIR v2 conformance tests

Sole owner of the classification and policy engine inside Ocular, an enterprise data governance catalog. No other engineer worked in the layer.

• Built a three-layer hybrid PII detector on Presidio, GLiNER, and transformer NER
• Shipped the policy platform so governance teams control scans instead of filing tickets
• Wired the first end-to-end path: Dremio to Gitea to Dgraph, over FastAPI and Airflow

A pre-registered study on whether MCP tool surfaces actually make LLM agents better, with the null result committed in advance. Aanish designed and ran it; I was one of two named approvers on the pre-answer anchors.

• The protocol bars the experiment host from approving its own anchor, so each sealed commitment went through outside review before any model call
• Question-only context selection beat query-blind projection by 9.5 points, on 43% less payload
• Replicated on an untouched 374-question holdout: +34.1 points on the microbiology stratum

Data-driven symptom tracking platform for Tourette's patients. Won 1st place at IIM Ahmedabad Tech for Social Good Hackathon.

• Pitched to 3-judge panel against 50+ teams nationwide
• Identified $5.6B addressable market opportunity
• Built prototype enabling real-time symptom tracking

ML-driven portfolio optimization improving risk-adjusted returns during high-volatility regimes. Supervised by Prof. Sandeep Juneja (PhD Stanford).

• Improved Sharpe ratio by 9%, Sortino ratio by 24%
• Trained diffusion models on 13+ years of NSE data
• Validated through 157 out-of-sample folds

End-to-end ML pipeline for billable hours forecasting serving consulting, legal, and nonprofit organizations.

• Built AWS serverless solution processing 100,000+ records
• Achieved MAPE <12% across 3 demo clients
• Evaluated 8 time series models for production deployment

Real-time event alerting system using SignalR for wealth management platform at Edelweiss Group.

• Served 1,000+ users across 12 modules
• Designed schemas for 25+ event types
• Captured 500,000+ events daily with 99.95% reliability