01 / AGENTS
Context-aware AI
Conversational and multilingual assistants with thoughtful prompts, memory handling, voice interfaces, and real task orientation.
Sydney, Australia · Remote-first
Build AI that
earns trust.
I evaluate, improve, and build AI systems for real workflows — from LLM quality and safety to practical agents, automation, and model benchmarks.
Signal layerEvaluate · refine · shipPractical intelligence
About
A builder’s mindset.
An evaluator’s rigour.
I’m Rownak Dip, an AI quality and applied AI specialist based in Sydney. I work at the point where useful AI meets reliable AI.
My experience includes evaluating production language and search-model outputs against detailed quality and safety rubrics, building conversational and multilingual AI tools in Python, and researching Vision-Language-Action robotics at Macquarie University.
I enjoy turning complex systems into something people can understand, test, and use with confidence.
26×faster VLA inference achieved through architecture benchmarking and efficient fine-tuning
LLM QAProduction evaluation across accuracy, usefulness, safety, and edge cases
PythonAgents, APIs, generative media workflows, and backend foundations
100%hands-on learning mindset — from research to production AI evaluation
Capabilities
Practical intelligence,
built to move.
I focus on the complete loop: build the useful thing, test what can go wrong, and make decisions from evidence rather than hype.
02 / EVALUATION
Quality that holds up
Structured model evaluation for accuracy, relevance, safety, hallucinations, calibration, and the edge cases that shape trust.
03 / AUTOMATION
Workflows with intent
Python and API-driven systems that link tools, reduce repetitive work, and surface the information needed for better decisions.
Selected work
Evidence, not
empty claims.
A few outcomes that represent how I work: investigate carefully, document what matters, and improve the system.
Teaching robots from video demonstrations
Macquarie University research internship using a mobile dual-arm ALOHA robot. Prepared RLDS TFRecords, benchmarked VLA architectures, and used LoRA/PEFT with OpenVLA.
~26× faster inferenceHigher task performance under a 98 GB VRAM constraint through model selection and fine-tuning.
Production AI model evaluation
Assessing language and search outputs for factual accuracy, relevance, usefulness, policy compliance, safety, and subtle failure modes.
Calibration-led qualityEdge-case documentation and structured feedback strengthen repeatable rating standards.
Conversational and generative AI experiments
Designed personal assistant, accountability, and multilingual agent workflows with LLM APIs, TTS/STT, and prompt/memory management.
Measured trade-offsCompared tools and models on response quality, API cost, latency, and practical suitability.
How I work
Build with intent.
Test with rigour.
01
Understand the workflow
Start with the real user, bottleneck, risk, and outcome that must improve.
02
Forge the useful loop
Design the smallest agent, workflow, or evaluation harness that creates visible value.
03
Measure and refine
Test behaviour, document failure cases, compare alternatives, and expand only what works.
Open to opportunities
Let’s make AI more
useful and trustworthy.
I’m open to remote-first AI quality, model evaluation, prompting, applied AI, and collaborative technical projects.