Abhinav (August) Srivastava

An open parser and an empirical study of VistA

• Open tree-sitter grammar for MUMPS, the language underlying the U.S. Department of Veterans Affairs hospital system, producing a reusable and analyzable concrete syntax tree where no public machine-readable grammar previously existed.
• Parses 98.6% of VistA's 33,951 routines (3.6M lines) with no error nodes; expression semantics verified by differential testing against YottaDB.
• Call-graph and runtime-code-generation measurements over the OSEHRA VistA-M corpus.
• doi:10.5281/zenodo.20782673

FHIR-native time-series storage for critical care

• Time-series database treating FHIR Observations as first-class citizens for ICU monitoring, eliminating the impedance mismatch between clinical data models and storage.
• Chunked storage with FHIR-aware metric encoding; integrated multivariate pattern detection covering seasonal decomposition, change-point detection, and anomaly scoring.
• Benchmarked against InfluxDB, TimescaleDB, and QuestDB on MIMIC-III (330M observations, 38K patients).

Adversarial evaluation of AI systems handling PHI

• Security evaluation of commercial and open-source voice AI platforms deployed in clinical settings.
• Testing framework built around a FHIR-compliant EHR system for realistic attack simulation, covering authentication bypass, prompt injection, PHI extraction, and synthetic voice attacks.
• Proposes a tiered security framework that reduces attack success rates while preserving clinical workflow.

Secure multiparty computation for cross-institutional ML

• Custom SMPC protocol enabling joint model training across hospitals, insurers, and public health agencies without pooling raw data.
• Additive secret sharing with Beaver-triple-based multiplication under an honest-majority assumption, plus a commitment-based verification layer for detecting malicious participants.
• Optimised for gradient-based workloads, requiring fewer communication rounds than general-purpose frameworks.

Zero-knowledge proofs for negative STR DNA match verification

• Interactive zero-knowledge protocol proving DNA non-match without revealing genetic data, operating over committed STR allele values with locus-level inequality proofs.
• Formal security analysis establishing completeness, soundness, and the zero-knowledge property.
• Targeted at forensic investigation, where a suspect requires privacy-preserving exclusion.

Lightweight formal methods for EHR data models

• Framework applying formal verification to the HL7 FHIR Patient resource, formalising structure, cardinality constraints, choice-type invariants, and cross-field dependencies in Alloy.
• Resource-centric verification boundary modelling external references as opaque types to keep analysis tractable.
• Bounded model checking to exhaustively verify structural integrity, semantic consistency, and safety invariants; compositionally extensible to other resources and cross-resource workflows.

Reproducible simulation for benchmarking clinical decision policies

• Emergency-department simulator modelling intake, triage, resource scarcity, and inventory under a randomised but clinically plausible event schedule.
• Every event logged as JSONL; a given scenario, seed, and action sequence replays exactly, so two policies can be compared under identical conditions.
• Exposed over MCP so language-model agents can be benchmarked as decision policies against a fixed environment.
• github.com/AbhinavMir/hospital-gym

Revenue-cycle automation for emergency medical services — San Francisco / remote

• Own all engineering — architecture, delivery, infrastructure, and roadmap — and lead two engineers and one intern.
• Architected and operate a claims pipeline processing $1M+ per month across 100+ agencies: ingestion, automated denial detection and classification, and resolution routing over X12 276/277, with adapters for ten or more payer, clearinghouse, and EMS systems, each implementing the EDI standard differently.
• Led the organisation through HIPAA and SOC 2 from zero: control design, evidence collection, and audit readiness.

Conversational-AI platform for medical practices — remote

• Built core platform infrastructure; grew the engineering team from one to three while continuing to ship.
• Delivered insurance verification, FHIR integration, robotic process automation, outbound calling, and a whitelabel interface, adopted by four practices in three months.
• Established systematic model benchmarking, prompt iteration, and observability to bring live patient-facing conversation quality to production standard.

Ambient clinical documentation used by thousands of clinicians — Mountain View, CA

• Integrated 15+ electronic health record platforms (athenahealth, eClinicalWorks, NextGen, DrChrono, ModMed, Elation, and others), each with distinct identity models, data schemas, and write-back semantics.
• Owned the clinician-facing mobile application and browser extension; maintained production infrastructure and databases for the flagship product.