Dinesh Pushparajah.

Over twenty years of shipping software: e-commerce and content platforms in the early web, a decade running engineering for digital agencies, then co-founding an AI startup and leading it as CTO.

Along the way I scaled a SOC 2 Type II platform from zero to 84,000+ users, directed a funded applied-AI research program, hired and mentored engineering teams, and most recently led a small team shipping a React Native app and Node.js backend daily.

Fine-tuning & distillation

Tuned frontier models for narrow jobs, and distilled them small enough to run on-device over WebGPU with sub-second, private inference.

Retrieval-augmented generation

Vector search with adaptive semantic chunking and cross-lingual English/French embedding alignment over messy, mixed document sets.

Vector embeddings & matching

Vertex AI embeddings with pgvector similarity search, plus an LLM reasoning pass that scores every match before a person sees it.

Agent workflows

Autonomous agents that explore and structure web pages they have never seen, and multi-model chains that extract constraints, classify intent, and validate grounding.

Guardrails & streaming

Fail-closed prompt-injection guards and real-time streamed generation over SSE, shipped on a SOC 2 Type II platform.

Multi-provider routing

A provider-agnostic gateway across Gemini, Claude, and OpenAI with rate limiting, so switching models never needed a code change.

Predictive modelling

Temporal-aware scoring models with fairness calibration on skewed historical data, and computer vision that reconciles what is on screen with the DOM.

Trained through a national AI accelerator, and I write software with AI every day: Claude Code, Cursor, and Codex are part of the workflow, not an experiment.

The whole stack for this page is the stack I work in. I’d rather run it than describe it.