TFX custom schema component
Built a schema component for TFX-Addons, the community extension library for TensorFlow Extended pipelines, through code review with maintainers.
AI INFRASTRUCTURE
The unglamorous part of AI: serving it reliably.
Models are only useful when they're fast, observable and affordable in production. I bring Kubernetes and SRE experience to model serving, plus open-source ML pipeline work and a research background.
01 / FEATURED PROJECT
In progressCrucible: a self-hosted LLM inference gateway on Kubernetes.
An OpenAI-compatible gateway in front of self-hosted models, built for the problems teams hit once an LLM feature has real traffic: cost, latency, fairness between tenants, and safe model upgrades.
Client · OpenAI SDK
Go gateway
Router · canary
Model servers
Redis · token rate limits
Postgres · usage & keys
Prometheus · TTFT, tokens/s
KEDA · queue-based scaling
Streaming proxy
OpenAI-compatible API with token streaming, so existing SDKs work unchanged.
Token-aware limits
Per-key budgets measured in tokens, not requests, with usage accounting.
Scale on real load
Autoscaling on queue depth and in-flight requests instead of CPU.
Safe model upgrades
Canary routing between model versions, gated by an eval suite in CI.
02 / ML PIPELINES
Open-source ML infrastructure.
Metrics into analytics
Containerized microservices on Kubernetes and connected them to Oracle Analytics Cloud for metrics visibility.
03 / RESEARCH
Comfortable with formal depth.
MS by Research at IIIT Hyderabad's Centre for Quantum Science and Technology. I can read papers, reason about proofs, and turn ideas into working systems.
Maximal Secret Reconstruction, Teleportation and Bell's Inequality ↗Full research page →04 / HOW I CAN HELP
Self-hosted model serving
Run open models on your own infrastructure, with autoscaling and cost visibility.
LLM app observability
Latency, token usage and cost per feature, on dashboards your team will actually use.
Productionizing ML pipelines
Move notebooks and scripts into containerized, scheduled, monitored pipelines.
Shipping an AI feature that needs to survive real traffic?
Freelance projects, long-term contracts, or full-time remote AI infrastructure and MLOps roles.