10x throughput · Kafka · 13 social channels · built for viral-scale message bursts
Redesigned an enterprise social platform's data layer into an event-driven architecture on Apache Kafka, ingesting and analysing real-time social data from 13 channels.
Sized for viral-scale bursts of 250–1,000 messages/minute per customer, with hybrid delivery paths: streaming (~15–20s average latency) and polling (~2 min), with a worst-case 5-minute latency ceiling.
Also set architecture standards across application, data, and AI, and rebuilt the React/TypeScript frontend for multilingual workspaces with automated WCAG 2.2 / ADA Section 508 accessibility checks. End-to-end throughput improved 10x. Confidential client engagement.
0
24
Cold start: 17 min → under 3 min. Shard loading: 16 min → 50 sec. Serving cost: ~1/6 of a managed SageMaker equivalent.
Performance-engineered a LLaVA vision-language deployment on AWS as part of a specialised strike team, with the full write-up published on LinkedIn (linked under Completed Work).
Reduced end-to-end cold start from ~17 minutes to under 3 (66%) via custom Deep Learning AMIs (−32% immediately), ephemeral-storage model caching with parallel aria2c downloads (checkpoint loading 16 min → 50 sec), and 4-bit NF4 quantization (shard loading 41s → 18s). Along the way, diagnosed a storage-path bug where checkpoint shards kept re-downloading despite being baked into the AMI.
0
49
Cut shipping time 40% with a production multi-agent system built around governance, not demos. Confidential client engagement.
Designed and deployed a multi-agent system covering the full product lifecycle: PRD generation → system design → multi-repo execution → code review → compliance/security checks → release management. Shipping velocity improved 40% across cross-functional teams.
Built on a Backstage-based engineering knowledge base with MCP connectors to JIRA, GitHub, Slack, and Teams; LLM workloads on OpenAI and Anthropic via A2A-style agent orchestration, implemented in Python/Bash tooling.
Shipped with an enterprise AI governance layer: hallucination detection with knowledge-base remediation, model drift monitoring, and audit logging for regulated environments.
0
50
90M registered users · 17 countries · 99.99% uptime · $451M disbursed via DBT
Led Everwell Hub from a monolith to a horizontally scalable microservices platform (Java/Python, Docker/Kubernetes, GCP, CI/CD, IaC). As chief architect I personally designed the service decomposition; as CTO I led ~50 engineers plus system-integrator partners and the WHO/Ministry of Health partnership.
Deployed across 17 countries and scaled India's national TB platform (Ni-kshay) to ~90M registered patients and staff, with 6 lakh+ direct users, at 99.99% uptime, sub-second P95 latency, and 2h RTO / 4h RPO disaster recovery. The payments module disbursed $451M through direct benefit transfers.
Released the platform as MIT open source with full wiki, runbooks, and API contracts — adopted by partner health systems since.
0
76