Freelance solutions architect · 16 years
I design distributed systems that stay up — and ship the agentic products that run on them.
Sixteen years across healthcare and financial services, most recently as Technical Architect and Engineering Manager at AthenaHealth: a platform serving 50M+ requests a day at 99.99% uptime, an ECS-to-EKS migration run hands-on, and clinical interoperability with Cerner, CommonWell and CareQuality over HL7 FHIR.
I now take that work on directly — architecture, agentic systems, and end-to-end delivery for teams that need the thing built, not just diagrammed.
Requests per day on the platforms I architected
Sustained uptime, fault-tolerant by design
Cycle-time reduction from AI, spec-driven delivery
Engineers across 7 teams on that workflow
Selected work
All eight projects →A freight load board for mid-size Indian carriers — public marketplace, broker console and a matching service that ranks loads against a truck's route, capacity and return leg.
An orchestrator for long-running agent workflows: durable state, typed tool contracts, compensation on failure, and human approval steps that survive a redeploy mid-run.
A clinical summarisation engine that reads a patient's scattered record — problems, allergies, notes, external documents — and hands the physician a structured, cited summary before the appointment starts.
A reliability and cost console for EKS: which workloads are over-provisioned, which are one node failure from an outage, and what each namespace actually costs per week.
Agentic AI
Your agentic journey, taken to production
Most agent projects stall between the demo and the deployment: no evaluation harness, no cost ceiling, no idea what happens when a tool call fails at 2am. That gap is an architecture problem, and it is the one I solve.
How I work on agents →Tool boundaries, state, retries and human checkpoints designed as a system — not a prompt chain.
Golden sets, regression gates in CI, refusal and escalation paths, PII handling for regulated data.
Model routing, caching, batching and token budgets, with per-tenant spend visible in Datadog.
The workflow that cut cycle time 60% across 35 engineers, installed in your team's own repos.
Puneet arrived with a diagram and left us with a running platform. Six months in, the payments core has not had a Sev-1.
Have a system to build?
Thirty minutes, no deck. Bring the problem and we will scope it on the call.