AI ENGINEER · SYDNEY, AUSTRALIA

Bisar UlHasan

I build agentic and generative AI systems, and I help companies and people put AI to real work.

Get in touchSee my work
Focus
Agentic + GenAI
Stack
LLMs · RAG · Agents
Now
Building daily
LLMs/RAG Pipelines/Multi-agent/LangGraph/LangChain/MCP/Vector Search/Fine-tuning/FastAPI/PyTorch/Evaluation/Prompt Engineering/Vercel/Supabase/Docker/Python/LLMs/RAG Pipelines/Multi-agent/LangGraph/LangChain/MCP/Vector Search/Fine-tuning/FastAPI/PyTorch/Evaluation/Prompt Engineering/Vercel/Supabase/Docker/Python/

01 /  About

I turn repetitive, slow work into AI systems that ship.

Growth-orientedSolution-firstShips fastGoes deep on the detail

I am an AI Engineer based in Sydney, working across agentic and generative AI: LLMs, multi-agent pipelines, RAG systems, and the data pipelines underneath them. My day job is driving AI adoption and automation inside a school, where the work is less about demos and more about systems people rely on every day.

Before this I helped build a speech-AI startup from a small founding team into a full company, leading dataset and model work for production text-to-speech. Earlier still, I did multimedia engineering for large institutional clients. Different tools each time, same instinct: find the slow, manual thing and build something that does it better.

I have been a gamer since 2009, and that is where the working style comes from. I stay in a problem until it is solved, and I go deep on the detail rather than moving on before I understand the failure. I default to what is possible, I pick things up fast, and I am usually shipping something new.

02 /  Work

Real systems, shipped.

Problem, build, outcome — no fluff.

RAG / PRODUCTION01

Teaching Assistant Bot

Problem
Staff needed fast, trustworthy answers from a large, messy body of school documents, without sending anything to a third-party cloud.
Built
A production RAG system with hybrid retrieval, cross-encoder reranking, citation enforcement, and a CI-gated evaluation pipeline so quality is checked on every change. Runs locally on open-source models with a cloud fallback.
Outcome
Source-cited answers people can trust, with retrieval quality measured on a golden evaluation set rather than guessed at.
PythonLangChainWeaviateOllamaCohereRAGASGitHub Actions
View on GitHub
DOCUMENT AI02

Teaching Program Automation

Problem
Writing teaching programs across the school's full subject list was slow, repetitive, and ate weeks of staff time.
Built
An AI document-processing pipeline that drafts programs from source material and structures them to the required format, leaving staff to review rather than write from scratch.
Outcome
Turned a multi-week manual job into a review-only workflow, across the whole subject list.
PythonPlaywrightLLMs
MULTI-AGENT03

Meeting Minutes Pipeline

Problem
Turning raw meeting transcripts into clean, formatted minutes by hand was tedious and easy to get behind on.
Built
A three-agent pipeline — transcript processor, minutes writer using Anthropic tool calling, and document generator — exposed via a FastAPI webhook with async background execution.
Outcome
Raw transcripts become formatted Word minutes automatically, with a person only in the loop to approve.
PythonFastAPIAnthropicMulti-agent
DATA PRODUCT04

NAPLAN Analytics Platform

Problem
Raw national-assessment results were hard to turn into decisions, and analysing them by hand took weeks.
Built
A web platform that turns raw assessment exports into interactive dashboards for school-wide decisions, plus personalised practice that targets each student's weakest skills.
Outcome
Collapsed weeks of manual analysis into a near-instant view, and made the data usable by people who are not analysts.
ReactVercelSupabasePostgreSQLEdge Functions

Supporting experience

SPEECH AI / SUPPORTING05

Production TTS at Scribe Audio

Problem
A speech-AI startup needed production-quality text-to-speech and the data pipeline to train it, built from the ground up.
Built
Led dataset creation with professional voice actors and processed thousands of hours of audio into ML-ready data. Trained and fine-tuned Tacotron 2 and neural vocoders for production speech synthesis, and cut model training cycles sharply by adopting better architectures.
Outcome
Production TTS deployed for enterprise audiobook clients. Grew with the company from a small founding team to a full organisation, leading a team of engineers.
PyTorchTacotron 2WaveGlowLibrosaAWS
ENGINEERING / SUPPORTING06

Berserk Media

Problem
Enterprise and institutional clients needed web, digital, analytics, and media work delivered to a high bar.
Built
Multimedia and web engineering for clients including the World Bank, British Council, US Embassy Pakistan, and NADRA. Progressed from intern to specialist.
Outcome
The engineering and delivery foundation the AI work was later built on.
WebMultimediaAnalytics
berserk.com.pk

03 /  AI focus

How I think about AI adoption

01

Start from the boring, repetitive work

The best AI wins are not flashy. They are the slow, manual tasks people do every week. Find one, automate it well, and the value is obvious.

02

Measure it or it does not count

An AI system you cannot evaluate is a demo. Build the evaluation set, gate quality in CI, and you can ship changes without crossing your fingers.

03

Keep a person in the loop where it matters

Automation should draft and propose. People approve. That split is what makes AI safe to put in front of a real organisation.

Writing

Follow on LinkedIn →
AI FUNDAMENTALS

[FILL IN] — agent patterns explainer

COMING SOON
BUILD WITH AI

[FILL IN] — shipping with Claude Code

COMING SOON
AI USE CASES

[FILL IN] — a real automation case study

COMING SOON
PRODUCTIVITY

[FILL IN] — the AI tool stack I actually use

COMING SOON
DATA + NEW TECH

[FILL IN] — sharp take on a new model

COMING SOON
GROWTH

[FILL IN] — learning AI without drowning

COMING SOON

04 /  Contact

Let's build something with AI.

Whether you want to adopt AI in your business, ship a system, or just compare notes, my inbox is open.

bisar1000@gmail.com
LinkedInin/bisarhasanGitHubbisarulhasanX[FILL IN]