Evidence is not accepted truth
Source records remain immutable until an authorized reviewer accepts a proposed change.
From messy workflows to reliable AI products.
I take ambiguous user problems from discovery to shipped systems, owning the product decisions, architecture, evaluation and delivery.
Sydney, Australia · Building across product discovery, agentic systems, retrieval, evaluation and full-stack delivery.
OrchestraA product brain for client-facing software teams.Synthetic demonstration data
100+customer interviewsOrchestra discovery
5pilot customersOrchestra beta
17read-only integrationsOrchestra context layer
148evaluation casesretrieval and grounding suite
Orchestra
A product brain for client-facing software teams.
The problem
Product decisions fracture across meetings, documents, conversations and code. Teams repeat discovery, lose rationale and act without shared context.
The product
Orchestra separates immutable project evidence from reviewed product truth. It turns PRDs, meetings, client conversations and code into citation-backed answers, proposed requirement changes and approved context that can be delivered to teams and AI coding tools through MCP.
- Ownership
- Discovery, strategy, architecture, evaluation and delivery
- Status
- Private beta
- Evidence
- 100+ interviews, five pilots, 148 evaluation cases
Product memory architecture
- Sources
PRDs, meetings, conversations, GitHub and Drive
- Immutable evidence
Source records preserved with provenance and permissions
- Hybrid retrieval
Dense and sparse search with metadata filters and reranking
- Review proposal
Questions, conflicts and requirement changes mapped for review
- Versioned truth
Accepted changes update auditable Product Brain state
- Socrates and MCP
Grounded answers and approved context delivered where work happens
Review precedes version change
Live Doc proposals map changes to product context before Product Brain state moves forward.
Evaluate before expanding
A stable suite protects relevance, grounding and regressions.
Selected work
Four more systems spanning agent orchestration, operational automation, analytics and evidence-first knowledge work.
02Public open-source release
LaneOrchestrator
A policy-aware orchestration layer that routes Codex work across model lanes, specialist agents, approval gates and verification workflows.
172 specialists · 3 model lanes · zero third-party runtime dependencies · reviewed v0.2.4
Ownership: Product concept, Routing architecture, Policy design
View case study
03Live synthetic demonstration
ReplayOS
A workflow control plane that converts operational evidence into typed, approval-gated execution with recovery and verifiable traces.
AUD 500 approval threshold · exact AUD 740 effect binding · one idempotent effect · versioned recovery
Ownership: Product strategy, Workflow model, Control architecture
View case study
04Open-source reference build
Governed Analytics Copilot
A governed analytics copilot for natural-language SQL, forecasting, approvals and evidence-backed business decisions.
SHA-256 query and dataset binding · 5 second execution budget · 1,000 row cap · deterministic offline path
Ownership: Product framing, Governance architecture, Backend and frontend delivery
View case study
05Open-source reference build
Evidence-first Knowledge Workspace
A role-aware knowledge workspace that turns approved documents into grounded answers with evidence, confidence signals and human review.
4 role surfaces · hybrid retrieval with reranking · citation cards · confidence-gated review
Ownership: System design, Agent workflow, Retrieval and evaluation
View case studyHow I work
One operating model, from first conversation to verified product behaviour.
- 01
Find the real decision
Start with the user, the workflow and the cost of getting the decision wrong.
- 02
Shape the smallest useful product
Turn the problem into a focused journey with a clear result and measurable behaviour.
- 03
Design the authority boundary
Decide what models may propose, what software must enforce and where people stay in control.
- 04
Build and evaluate
Implement the complete path and test retrieval, grounding, permissions, failure and cost separately.
- 05
Ship the operating system around it
Treat review, observability, recovery and evidence as product surfaces, not release notes.
Experience
Work across product discovery, applied AI and full-stack delivery.
Founder / AI Builder in Residence
Arrayah and SH1P AustraliaTook Orchestra from 100+ customer interviews to a live beta with five pilot customers, owning discovery, system architecture, backend delivery, evaluation and pilot execution.
AI and Automation Engineer
KRSP TechCreated a privacy-first candidate-ranking system with section-aware scoring, local document processing and auditable LLM-assisted explanations.
Artificial Intelligence Intern
Aavaaz IncDeveloped NLP and speech-recognition pipelines for real-time translation, reaching 97% language-detection accuracy.
Artificial Intelligence Intern
Continental AutomotiveWorked across real-time terrain perception, an internal RAG assistant adopted by 3,000+ employees and Python automation that reduced documentation effort by roughly 95%.
About
I do my best work where the problem is still unclear. I enjoy talking to users, mapping the workflow, deciding where models should and should not have authority, and then building and evaluating the system end to end.
Education
Master of Data Science and InnovationUniversity of Technology Sydney2025 to 2027
Bachelor of EngineeringBMS College of Engineering2020 to 2024
Recognition
- Selected for Arrayah Accelerator Chapters 3 and 4
- Selected for SH1P Australia Cohort 1
- Showcased an AI-enhanced terrain-adaptive vehicle control system at Continental Innovation Day
- Submitted the terrain-adaptive vehicle control system to SAEINDIA International Mobility Conference 2024
Product discovery and strategy
Customer discovery, 0 to 1 product development, requirements shaping, workflow design, rapid prototyping, evaluation design and pilot delivery
Agentic systems and retrieval
Agent orchestration, hybrid retrieval, reranking, embeddings, vector search, tool calling, MCP, grounding and citation verification
Full-stack delivery
Python, TypeScript, SQL, FastAPI, Node.js, React, Next.js, PostgreSQL, pgvector, Redis, Prisma and BullMQ
Evaluation and reliability
LLM evaluation, regression testing, CI/CD, RBAC, audit logs, observability, sandboxing, security scanning and workflow automation
Applied machine learning
NLP, computer vision, speech recognition, PyTorch, Hugging Face, scikit-learn, Pandas, NumPy and data visualisation
Building something where AI has to work inside a real workflow?
I’m interested in teams that care equally about user value, technical depth and what happens after the demo.
Sydney, Australia · Open to ambitious product and engineering teams.