Adi Prabs
Computing @ Imperial. SRE @ Apple. I build production AI systems on the side — compilers, infra, agents, and the boring glue that turns demos into products.
- —Apple — SRE on ML Platforms.
- —Side work with funded startups (ARR-stage).
- —Exploring next venture: physical AI / edge AI / hardware-software plays.
- —Project Nine: a climbing quadruped robot, MuJoCo-first.
Where I've shipped.
Six companies, three years. Each one taught me something specific about turning research into production — latency, cost, reliability, security, or all at once.
Apple
Site Reliability Engineer, ML Platforms- —Built a Kubernetes capacity forensics platform end-to-end (collectors, scanners, delta-query UI) used by 30+ SREs to diagnose EC2 capacity exhaustion, saving up to $3M per AWS availability zone annually.
- —Built a cloud-agnostic capacity request and reservation management system — replaced ad-hoc Slack coordination with an auditable workflow handling hundreds of requests monthly.
- —Created a Kubernetes manifest validation framework detecting misconfigurations at deploy time; retrospective analysis shows it would have caught 72% of deployment-related incidents over the prior year.
8x
Full-stack & AI/ML Developer- —Optimized production analytics from 24s to sub-second via SQL-side aggregation and indexed Postgres RPC rewrites; cut /posts payloads 90%+ (22MB → ~1–2MB).
- —Simplified messaging architecture, deleting ~400 lines of legacy API code while enabling a new admin reply UX.
- —Resolved 3 critical production vulnerabilities: org takeover, exposed financial Server Actions, and DB search-path injection across 49 functions.
Canopy Labs
General Engineer- —Architected, built, and deployed the company's flagship full-stack web application, enabling real-time multi-user usage with Docker and Kubernetes orchestration.
- —Created agents to enable custom form filling from transcripts, cutting insurance resolution by 15 min per form.
- —Optimized backend concurrency and load balancing, reducing latency 500ms → 120ms, supporting 100+ active users.
- —Implemented AWS CI/CD pipelines with automated testing, shortening release cadence to 6 hours.
Vani
Full-stack & AI/ML Developer- —Architected, shipped, and deployed the flagship multi-tenant web app on AWS with Docker + Kubernetes.
- —Cut p95 backend latency 500ms → 120ms; scaled to 100+ concurrent users.
- —Stood up CI/CD: release cadence 2 days → 6 hours, production bugs −76%.
- —Integrated LLM workflows via Model Context Protocol for clinical-admin automation.
- —Held 99.9% uptime through staged rollouts and load-balanced workers.
Trajex
Machine Learning Developer- —Deployed LLama 3.2-7B-Instruct in production — 20% cost reduction vs OpenAI, 12% lower inference latency.
- —Led product design and built the inference backend.
- —Pitched investors and onboarded K3 Capital Group as a paying client.
Altus Reach
ML Engineer (Contract)- —Team of 3 — built a video saliency model improving prediction accuracy by 19%.
- —Shipped Azure-hosted inference pipeline for production traffic.
- —Full-stack work on company web app (TypeScript / Next.js / React).
Things I've built.
all projects →I'm Adi — Computing student at Imperial College London, currently SRE on the ML Platforms team at Apple.
On the side, I ship production AI systems for early-stage startups. Past lives: healthcare admin automation (Vani), LLM cost-reduction at Trajex, video saliency ML at Altus Reach.
I like systems that work at small scale and don't fall over at large scale — compilers, infra, agents, and the boring glue that turns demos into products.
Reachable at adiprabs19@gmail.com — most useful when there's a concrete problem attached.
Got a hard problem worth solving?
I take on a small amount of side work and I'm always interested in talking to people building something genuinely difficult — especially in physical AI, dev infra, or edge inference.