- Lead developer · Nuthan TechnologiesEnterprise healthcare analytics platform7 production modules shipped
- Builder · Nuthan TechnologiesWarehouse migration MCP tool2–3 days of manual scripting replaced per migration
- Co-founder & IT Lead · side ventureChartHaven0 patient records stored on the server
- Georgetown MHSA capstoneAI hospital routingd = 0.64 effect size vs closest-hospital routing
- Personal projectHermes Agent8 min from a red build to a green fix, untouched
- Personal projectSelf-hosted homelab5 machines watched on a private mesh VPN
Nirvahn ThakurMHSA
I help organizations implement AI properly, and I'm also the engineer who builds it.
Enterprise Architect @ Workday · Charlotte, NC
Systems I've built and run
Press and hold. Every slip opens a case study.
Sprayed 00
7production modules shipped
Enterprise healthcare analytics platform
Lead developer on a production analytics platform, from live data to AI agents.
2–3 daysof manual scripting replaced per migration
Warehouse migration MCP tool
An MCP server that automates moving a data warehouse, starting with Snowflake to Databricks.
0patient records stored on the server
ChartHaven
A patient-owned medical records app whose AI answers only from your own records.
d = 0.64effect size vs closest-hospital routing
AI hospital routing
Patient routing that weighs specialty fit and ED crowding, not just distance.
8 minfrom a red build to a green fix, untouched
Hermes Agent
A personal AI assistant I text or talk to. It plans my to-do list, waits for my yes, and hands coding to Claude Code.
5machines watched on a private mesh VPN
Self-hosted homelab
A private, fully monitored homelab of five machines, running the services I depend on.
More builds
- Claude Code harnessMy agent-coding setup, run like a system: measured, audited and kept lean, with hooks against context rot and a live phone board of every session.
- Self-healing knowledge systemA synced, end-to-end encrypted notes vault with deterministic maintenance and two-way task sync.
- ARIA recovery monitoringA voice-first post-op recovery companion that streams sensor data and emits FHIR bundles for the EHR.
- Offline voice assistantA push-to-talk handheld that runs speech and the LLM entirely on my own network.
- Prediction-market botA Python bot that turns outside signals into a probability, then compares it with the market price.
EMS dispatch rebalancing
Shift coverage redesigned from six months of dispatch data, about 12% faster response across 3 stations.
01 — ABOUT
About me
I'm an Enterprise Architect at Workday, where I help organizations turn AI into working systems across HR, finance and operations.
Before that I was lead developer on a production healthcare analytics platform, shipping agents, MCP tooling and the data pipelines behind them.
I came to this through healthcare: EMT, then pre-med psychology, then a Georgetown MHSA, so I know the care pathways these systems have to fit.
- EMT
- Psychology
- MHSA
- AI engineering
- Enterprise architecture
Mentor with Big Brothers Big Sisters, 2016–2020. Languages: English, Spanish, Hindi.
- Education
Georgetown University
MS, Health Systems Administration (MHSA)
2024 – 2026
Clemson University
BS, Psychology (pre-med track), minor in Business Administration
2020 – 2024
NCOAE
EMT (BLS) certification
2022
02 — SKILLS
The stack
Hover to preview an element, click to pin it
AI / LLM13
Claude Code · 1
Daily driverPrimary development workflow, with 15 custom slash commands, hooks and multi-agent orchestration.
See it in use →Claude API · 2
ProductionAgents and apps built on the Anthropic API.
Azure OpenAI · 4
ProductionChat and extraction on GPT-4-class models behind production agent backends.
See it in use →Embeddings · 6
Productionbge-large retrieval that lifted recall@5 from 0.929 to 1.000.
See it in use →Model evaluation · 7
ProductionGold sets, F1 and decline-rate evals, and A/B comparisons that pick the model.
See it in use →Local models · 8
WorkingBenchmarked Llama, Mistral and Phi; fine-tuned and A/B-tested models on Apple silicon; Ollama at home.
See it in use →LoRA fine-tuning · 9
WorkingFine-tuned 7B/8B adapters locally, then retired them when a 14B base won the A/B.
See it in use →MLX serving · 10
WorkingSelf-hosted 14B model on Apple silicon with speculative decoding (about 1.8x faster).
See it in use →Code knowledge graphs · 11
ProductionCodebases as queryable graphs; agents read the graph report instead of raw files.
See it in use →Token cost engineering · 12
ProductionOutput compression, usage profiling and MCP schema pruning after a 97-session audit.
See it in use →Voice pipelines · 13
WorkingOffline Whisper speech-to-text and Piper text-to-speech over the Wyoming protocol.
See it in use →
Agents & MCP10
MCP · 14
ProductionA 10-tool migration server over Streamable HTTP, with the UI as an MCP client.
See it in use →LangGraph · 16
WorkingMulti-stage agent graphs with a deterministic verifier against hallucination.
Agent observability · 17
ProductionExecution tracing, failure monitoring and token-level cost tracking.
See it in use →Langfuse · 18
ProductionTraced every stage of agent pipelines in self-hosted deployments.
Multi-agent orchestration · 19
ProductionAgent teams and git-worktree-isolated workers with a completion relay.
See it in use →Claude Code hooks · 20
Daily driverHooks on 14 event types, including a per-turn canary against context rot.
See it in use →Pydantic · 21
WorkingStructured LLM outputs; certified.
AutoGen · 22
WorkingMulti-agent design patterns.
OpenRouter · 23
WorkingModel routing and fallback in Hermes Agent's first model chain.
See it in use →
Data12
Python · 24
Daily driverData pipelines, agents and backends.
SQL · 25
Daily driverPipelines and analytics queries.
Cortex Search · 28
ProductionManaged vector search on Snowflake for document RAG.
Supabase + RLS · 30
ProductionRow-level security on every table and an idempotent provisioning script.
See it in use →SQLite · 31
ProductionLocal-first stores: job queues in WAL mode, agent memory and trade logs.
See it in use →CouchDB sync · 33
ProductionSelf-hosted end-to-end encrypted notes sync across four devices.
See it in use →Knowledge pipelines · 34
ProductionDedupe-on-ingest, document-to-Markdown conversion and nightly health checks.
See it in use →Market data systems · 35
WorkingKalshi and Polymarket clients, macro feeds, half-Kelly sizing and a risk module.
See it in use →
Full-stack11
React · 38
ProductionFront ends for analytics and tooling.
TypeScript · 39
ProductionFront ends and this site.
Next.js · 40
ProductionFastAPI-backed web apps, and this site.
Tailwind · 41
ProductionDesign systems on Tailwind v4.
Auth & MFA · 45
Productionbcrypt, TOTP MFA, refresh tokens and least-privilege database roles.
See it in use →API integration · 46
ProductionTodoist, Gmail OAuth and the Obsidian REST API, with idempotent two-way sync.
See it in use →
Infra / DevOps13
Prometheus + Grafana · 50
Production15-second scrapes and generated, query-validated dashboards.
See it in use →Alerting & watchdogs · 51
ProductionTwo-strike alerts, a dead-man heartbeat and a stale-workflow detector.
See it in use →Linux networking · 52
ProductionRoot-caused a gateway fault with tcpdump and iptables tracing.
See it in use →CI/CD · 54
ProductionGitHub Actions, Dependabot, pre-commit with gitleaks, and an iOS build check.
See it in use →Testing · 55
Productionjest and supertest with a coverage gate; pytest regression suites.
See it in use →AppSec · 56
ProductionSemgrep SAST plus weekly ZAP, Nuclei and authorization-probe DAST.
See it in use →Home Assistant · 57
ProductionSmart-home hub on a Pi 5 with Matter devices and an MQTT bridge.
See it in use →Render / Vercel · 58
ProductionHosting for backends and this site.
Healthcare7
HIPAA-oriented design · 60
ProductionEncryption, audit logging and least-privilege access in ChartHaven.
See it in use →Applied cryptography · 61
ProductionSecure Enclave P-256, ECDH/HKDF and AES-256-GCM sealing.
See it in use →EMT (BLS) · 62
ProductionFrontline care at Vital Care EMS.
Interoperability · 63
WorkingFHIR R4 bundles with LOINC and ICD-10 coding; C-CDA record import.
See it in use →Health systems · 64
ProductionGeorgetown MHSA: care pathways and payer operations.
Decision modeling · 66
WorkingA utility function weighing severity, specialty fit, transport and ED delay.
See it in use →
04 — EXPERIENCE
Where I've worked
The path
- EMT
- Psychology
- MHSA
- AI engineering
- Enterprise architecture
- Technology
- Clinical
- Academic
Jul 2026 — Present
Enterprise Architect
Workday · Charlotte, NC (Technology)
- Help customer organizations design and deploy AI-driven solutions across HR, finance and operations.
- Translate platform and AI capabilities into clear strategy for business and technical stakeholders.
Jun 2024 — Jul 2026
AI Engineer
Nuthan Technologies · Charlotte, NC (Technology)
- Lead developer of an enterprise healthcare analytics platform: 7 production modules, including an AI Insights Agent and an agent-observability layer.
- Automated 3 manual ETL processes with LangChain and CrewAI agents, cutting analyst prep from about 4 hours to under 20 minutes per cycle.
- Shipped an MCP tool for cross-warehouse migration (Snowflake → Databricks), replacing 2–3 days of manual scripting per migration.
Feb 2023 — May 2024
Emergency Medical Technician
Vital Care EMS · Upstate, SC (Clinical)
- Analyzed 6 months of dispatch data to rebalance crew shift coverage, cutting average response time about 12% across 3 stations.
- Provided basic life support under time-critical conditions, with patient care records kept to state and federal standards.
Aug 2022 — Aug 2023
Academic Tutor, Organic Chemistry
Clemson University (Academic)
- CRLA Level 2 certified tutor.
Dec 2021 — Jan 2022
Student Intern
Atrium Health (Clinical)
- Clinical rounds in pediatric congenital heart disease.
Jun 2019 — Aug 2020
Intern
The Neurological Institute (Clinical)
- Supported clinical trials.
- Researched and presented on Myasthenia Gravis to the board of physicians.
Education
Georgetown University
MS, Health Systems Administration (MHSA)
2024 – 2026
Clemson University
BS, Psychology (pre-med track), minor in Business Administration
2020 – 2024
NCOAE
EMT (BLS) certification
2022
Certifications
- SnowPro Core (Snowflake)
- GitHub Foundations
- Claude Code in Action
- Developing Applications with LangChain
- Pydantic for LLM Workflows
- Multi AI Agent Systems with CrewAI
Also completed
- Building with the Claude API (Anthropic)
- Claude Code 101 (Anthropic)
- AI Fluency: Framework & Foundations (Anthropic)
- Introduction to Agent Skills (Anthropic)
- Introduction to Claude Cowork (Anthropic)
- AI Agentic Design Patterns with AutoGen (DeepLearning.AI)
05 — ACHIEVEMENTS
Proud moments
- 01
7
Production modules shipped on an enterprise healthcare analytics platform
2024 – 2026
- 02
28
Medical specialties of live HCP data behind one real-time backend
2024 – 2026
- 03
4h → 20m
Analyst prep per cycle, after agents automated three ETL processes
2024 – 2026
- 04
2–3 days
Of manual scripting replaced per warehouse migration by an MCP tool
2024 – 2026
- 05
8 min
From a red CI build to a green fix PR with nobody touching it: my agent planned, fixed and texted "Ready"
2026
How it works → - 06
9 s → 0.28 s
Memory recall per new chat for that agent, after moving memory to an always-on server
2026
How it works → - 07
2,710
Container restarts a ping-only monitor never saw in my homelab; health checks now catch it
2026
How it works → - 08
0% → ~98%
Task-sync success after I found a silent outage and added a relauncher that self-heals
2026
How it works → - 09
~12%
Faster average EMS response across 3 stations after shift rebalancing
2023 – 2024
- 10
d = 0.64
Effect size for AI hospital routing vs closest-hospital, across 500 simulated cases
2026
- 11
6
Certifications: SnowPro, GitHub, LangChain, CrewAI, Pydantic, Claude Code
2024 – 2026
- 12
MHSA
Master of Health Systems Administration, Georgetown University
May 2026
05 — ACHIEVEMENTS
Proud moments
06 — CONTACT
Let's build something.
Open to a conversation about enterprise AI, agents and healthcare systems. The fastest way to reach me is email.
Press ⌘K anywhere to jump around.
