Agentic AI engineering · Sydney
I'm Bisar Ul Hasan. I build the harnesses that run agentic AI in production: orchestration, memory, and LLM Ops.
Book my 30-min intro call
Founding engineer in a three-person speech-AI startup that scaled to 52 people. I led a team of nine and shipped production text-to-speech for enterprise audiobook clients.
Nine years on, my work is agentic AI harness infrastructure: the orchestration, memory, and LLM Ops layers that keep multi-agent systems running in production, with quality that is measured rather than guessed at.
Your agent demo already works.
Production is where it gets hard: state, retries, evals, cost, trust.
That gap is my job.
We pin down the highest-value thing to build and agree what good looks like.
I design and build the system end to end: data, models, evaluation, and the product around it.
It goes live with quality measured, and your team able to run it.
A service I built and run. Your scroll replays it.
A LangGraph state machine, run on real staff meetings.
Every run is traced, scored, and gated. Below baseline, it does not ship.
A person approves. Then it ships.
The meeting-minutes service from my agent operations platform — LangGraph over Claude on Amazon Bedrock, three-part memory on pgvector, Langfuse tracing, and eval gates that hold the line on quality.
A production agent is mostly the machinery around the model. Below, the machine walks itself one step at a time: the camera moves, each stage assembles, and the amber ball carries one real task through it, drafting the Year 12 Mathematics teaching program. The same pipeline covers 110+ subjects with about 80% less preparation time.
My own workshop runs on the same anatomy: an orchestrator routing a team of 17 specialist agents, with the same memory stores and ops loops underneath. I live in this stack.
A production RAG assistant over six textbooks. Hybrid BM25 and dense retrieval fused with RRF, Cohere reranking, and two-stage citation enforcement that declines ungrounded answers. Dual evaluation on a curated golden set (RAGAS faithfulness 0.94), gated in CI on every change.
Python · FastAPI · LangChain · Qdrant · Cohere · RAGAS · GitHub Actions
View on GitHub
A three-person speech-AI startup that scaled to 52. I led dataset creation with 50+ professional voice actors, processed 10,000+ hours of audio, trained the models that shipped to enterprise audiobook clients, and cut training cycles from 40 days to 5-10 by adopting better architectures.
PyTorch · Tacotron 2 · WaveGlow · Librosa · AWS
Also running: an AI document pipeline that automates teaching-program creation across 110+ subjects, cutting preparation time by about 80% and shifting the workflow to review-only.
It was an amazing experience to have Bisar as my team lead. He is an exceptional individual who can lead a team towards its goals — at Scribe Audio he introduced the team to the tools and technology stacks that accelerated our operations.”
He is not just a fantastic team leader and mentor, but an expert in his field who never stops learning — organized, practical, and innovative. When you need help with data operations, analysis, or any data-domain challenge, he is the one to turn to.”
It was an amazing experience working with Bisar. He is always active, energetic, and responsible, and never hesitates to take full ownership of the work. I highly recommend him as an engineer and as a team lead.”
Bisar is a great professional who has shown excellent capabilities in data engineering and curation. He is also great at managing operational teams.”
"I could give Bisar a problem and stop worrying about it. … Any team building serious AI products should talk to him."
30 minutes, straight into my calendar.
Your idea, where it stands, and the highest-value thing to build first.
If it fits, we scope the build. If it doesn't, that's fine too.
Prefer email? bisar1000@gmail.com