Software engineer · 4+ years · Charlotte, NC

I build healthcare and AI software, and I check that it works.

I take web applications from first commit to production: clinical data integrations, a HIPAA-compliant patient platform, multi-agent LLM pipelines and an open-source prior-authorization copilot on AWS. I own the backend, the AI features, the deployment and the monitoring.

Selected work

Three systems, start to finish

One open-source project I designed and run on my own budget, and two systems built for real users.

Featured · Personal open-source projectMIT licence · synthetic data only

Precertly

Open-source AI copilot for prior authorization: checks Medicare coverage criteria against a FHIR chart, cites a verified quote for every verdict, and always leaves the decision to a human reviewer.

Synthetic FHIR chart
Hybrid retrievalPostgres + pgvector
Rules layercoded data
Model judgmentnotes
Quote verification
Human reviewerdecides
96.2%criterion accuracy (177 of 184 criteria)
7 of 11cases correct overall
74.4%citation precision, target above 90%Not met

Experiment 1.1, as of 2026-10-28: 11 synthetic facet cases, one annotator, not clinically validated.

  • Python
  • FastAPI
  • PostgreSQL
  • pgvector
  • Amazon Bedrock
  • AWS CDK
George Mason University · 2025 – present

Carpal tunnel assessment platform

A browser app that lets diagnosed patients record hand movements at home and track recovery, while clinicians see structured results instead of raw video.

50+clinicians and patients in production
30%fewer clinician follow-up questions
90%test coverage, SonarQube-gated
1.5s→~0result delivery latency
  • FastAPI
  • PostgreSQL
  • MediaPipe JS
  • Celery
  • Redis
  • WebSockets
  • Azure Container Apps
Problem and what I built

The problem

Recovery from carpal tunnel is tracked in occasional clinic visits. Patients needed a way to measure progress between visits, and clinicians needed results they could trust without watching recordings.

What I built

  • Hand tracking in the browser (MediaPipe JS) feeding a four-stage signal-processing engine validated on 200+ patient recordings.
  • FastAPI, async SQLAlchemy and PostgreSQL backend: 40+ endpoints with JWT auth, per-patient RBAC and audit logging for HIPAA.
  • Celery and Redis workers, with results pushed over WebSockets the moment processing ends instead of 1.5-second polling.
  • A patient assistant that explains results in plain language, with PII masked before any model call.
Blue Horse Digital · 2024 – 2025

AI case generator

A platform that writes instructional case studies for training programs, grounded in client source material and graded before anyone reads them.

200+cases delivered each week
95%+of hallucinated content caught
800→<200retrieval latency, ms
25%lower Azure OpenAI spend
  • Next.js
  • FastAPI
  • CrewAI
  • LiteLLM
  • pgvector
  • Azure OpenAI
  • GitHub Actions
Problem and what I built

The problem

Writing cases by hand was slow, and early LLM drafts invented details. The client needed volume without losing accuracy, and leadership needed to know what each client's AI usage cost.

What I built

  • A CrewAI multi-agent pipeline, with LiteLLM routing each agent to the cheapest model that does its job well.
  • A retrieval layer on pgvector (HNSW plus re-ranking) that grounds every case in source material.
  • An evaluator agent with an eight-category rubric that flags fabricated content before delivery.
  • Per-run token and cost tracking, Azure Service Bus background jobs, and CI/CD with Cypress gates.
Approach

How I work

Habits I've built over four years of shipping, each backed by something you can check.

Own it end to end

From schema to API to UI to deployment, and the monitoring that tells me what I broke. I'd rather be the person who gets paged than the one who never hears about it.

  • GMUCarpal tunnel platform from browser hand tracking to Azure Container Apps.
  • PrecertlyEvery deployment run and verified from command output myself.

Measure before claiming

A number means nothing without what it was measured on. I publish the failing results next to the passing ones and write down what an experiment has to show before running it.

  • PrecertlyTwo quality gates reported as not met, front and centre.
  • PrecertlyA pre-registered experiment failed, so it was reverted and documented.

Tests are the gate, not the afterthought

Good tests save weekends. Nothing merges without green CI, and the test suite is designed alongside the feature.

  • GMUTDD with SonarQube quality gates, coverage raised to 90%.
  • IQVIARelease validation automated: regression testing from 3 days to 4 hours.
  • Precertly2,062 Python tests; pull requests merged one at a time after green CI.

Design to a budget

Cost is an architecture constraint. I compare designs with real numbers before choosing one, and track AI spend per run.

  • PrecertlyScale-to-zero AWS design under a $60 annual hard limit.
  • Blue HorseModel routing and cost tracking cut Azure OpenAI spend by 25%.

Keep humans in charge of high-stakes calls

In healthcare, software should make the evidence obvious, protect the data and leave the decision with a person.

  • PrecertlyNever denies a request; deterministic rules the model cannot override.
  • GMUPer-patient RBAC, audit logging and PII masking before any model call.
Experience

Where I've worked

May 2025 – Present

Software Engineer 2 · George Mason University · Remote

  • Lead engineer on the carpal tunnel assessment platform, from signal processing to production deployment on Azure Container Apps.
  • Introduced test-driven development with SonarQube quality gates in CI, raising coverage to 90%.
Jun 2024 – Present

Software Engineer · Blue Horse Digital · Remote

  • Built the AI case generator end to end, including the agent pipeline, RAG layer and cost tracking.
  • Standardized GitHub Actions CI/CD to Azure App Service, cutting deploys from 1 hour to under 10 minutes.
May 2021 – Dec 2022

Software Engineer · IQVIA · Bengaluru

  • Developed an event-driven Java and Spring Boot notification service moving 10K+ clinical record updates daily between EHR platforms and Salesforce.
  • Automated release validation for cross-system integrations, cutting regression testing from 3 days to 4 hours.
  • Built JUnit, Mockito and Postman suites for regulated clinical systems, reaching 85%+ coverage.
Jan 2023 – Dec 2024

M.S. Computer Science · George Mason University

Advanced Data Structures & Algorithms, Software Systems Design & Integration, Cloud Computing.

Toolkit

What I work with

Languages

  • Python
  • TypeScript
  • Java
  • SQL

Backend & data

  • FastAPI
  • Pydantic
  • SQLAlchemy
  • Alembic
  • Spring Boot
  • PostgreSQL
  • pgvector
  • Redis
  • Celery
  • WebSockets

Frontend

  • Next.js
  • React
  • Vite
  • Tailwind CSS

AI & retrieval

  • Amazon Bedrock
  • Azure OpenAI
  • CrewAI
  • LiteLLM
  • RAG
  • Hybrid search
  • Structured output
  • Evaluation design
  • MediaPipe JS

Cloud & delivery

  • AWS CDK
  • ECS Fargate
  • Aurora Serverless v2
  • API Gateway
  • Cognito
  • CloudFront
  • Azure Container Apps
  • App Service
  • AKS
  • Service Bus
  • Docker
  • GitHub Actions

Quality & security

  • TDD
  • Vitest
  • Playwright
  • Cypress
  • JUnit
  • SonarQube
  • CodeQL
  • HIPAA
  • RBAC
  • PKCE
  • TOTP 2FA
  • Audit logging

Healthcare standards

  • FHIR R4
  • HL7 Da Vinci PAS
  • Synthea
Contact

Hiring for full-stack, backend or AI engineering?

I'm open to full-time and contract roles across the US: onsite, hybrid or remote. Email is the fastest way to reach me.

tejavardhanr01@gmail.comSend an emailLinkedIn