Portfolio / AI engineering, 20266.9271° N, 79.8612° E
I build AI-powered software and the production systems underneath it: LLM applications, agents and agentic workflows that stay fast, measurable and recoverable under real load.
I build AI systems that stay measurable in production.
PL-01Three years of building real-time, multi-tenant systems on AWS taught me what actually breaks in production: latency, retries, cost and failure modes. I now point that instinct at AI, building LLM applications and agents that are bounded, measured and recoverable, not just impressive in a demo.
Features where the model is one component, not the architecture — provider abstraction, timeouts, validation and graceful fallbacks.
Tool-calling agents and multi-step graphs with explicit state, bounded loops and a human in the loop where the risk demands it.
Retrieval pipelines and schema-validated generation, so downstream code receives data it can trust.
Versioned prompts, tracing, per-tenant cost tracking and regression checks — measuring before optimising.
Event-driven, serverless and real-time infrastructure on AWS that the AI layer can lean on at scale.
Production systems and shipped products, written up the way I would explain them to another engineer: the problem, the architecture, and the decisions I made and rejected.
Product wanted AI summarisation. Engineering's question was how to ship it without making a slow, non-deterministic, swappable third-party API a load-bearing dependency of every request.
After the first LLM feature shipped, API spend spiked. Some tenants were heavy users, most barely touched it, and identical requests kept hitting the provider. There was no visibility into who was spending what.
AI-generated drafts were not accurate enough to ship directly, and one bad output in front of a client would damage trust. Reviewing everything by hand, though, would create a permanent backlog.
Single-model image generators put the burden of prompt-craft on the user. The goal was a product where a rough idea becomes a finished wallpaper without the user needing to know the prompt dialect.
A real-time chat platform serving 100K+ users had to stay reliable under bursty load, while one small team onboarded multiple client products onto the same codebase.
Agentic and LLM workloads are made of flaky, rate-limited async calls. Teams keep hand-rolling retry and concurrency logic around them, and getting it subtly wrong.
01 / AI
What I design
02 / Engineering
What I build it with
03 / Frameworks & platforms
What it runs on
Now — Current focus
Moving the systems instincts I built in production — latency, retries, cost, failure modes — into LLM applications, agents and evaluation. Building and writing up the patterns that make AI features dependable.
Apr 2023 — Present
Metarune Labs
Full-stack and distributed-systems engineer on a real-time platform serving 100K+ users with $10M+ revenue impact.
Feb — Apr 2023
Metarune Labs
Full-stack work across web, backend and mobile while learning the team's engineering standards.
Education / 2023 — 2027
SLIIT
Distributed Systems, Software Architecture, Application Frameworks, Data Structures & Algorithms.
Credential / Certified
Microsoft
Building AI applications and agents on Azure.
Verify credential ↗Experiments & side builds
Image and PDF tooling that runs entirely in the browser — WASM, Web Workers, zero uploads.
Next.js · WASM · pdf-lib
Booking platform for beauty and wellness businesses in a Turborepo monorepo.
NestJS · Prisma · Stripe
Construction project lifecycle, sustainability metrics and compliance documents.
Express · MongoDB · React
Building something where the AI part has to work in production? I am glad to talk about roles, projects and hard systems problems.