Learning from real corrections
Section-level edit metrics, anonymised before-and-after examples and user style profiles turn practitioner review into a controlled improvement loop.
Selected product work · six systems
Different industries. One ability: making complex products work.
Sombra leads because it proves I can deliver inside a regulated, high-stakes environment. The wider portfolio shows the range behind it—consumer products, generative media, private agent infrastructure, document intelligence and operational platforms.
Sombra
Building an adaptive AI operating system for a regulated industry.
Sombra turns fragmented meeting data and firm knowledge into reviewable file notes, client intelligence and compliance-aware workflows, then uses adviser corrections to improve future output.
The difficult problem
Financial advisers cannot safely connect client data to a general-purpose assistant and hope for the best. The system has to respect organisational boundaries, preserve an evidence trail, handle unreliable integrations and keep practitioners in control of sensitive work.
The product insight
Trust is not a disclaimer beside the AI. It has to live inside the workflow through visible sources, editable output, action boundaries, traceable activity and approval at the right moment.
What I built
- Meeting, recording and calendar ingestion with queues, retries and OAuth renewal
- Structured file-note and document workflows grounded in source material
- A feedback loop that measures edits and assembles quality-gated examples
- Production access controls plus broader agent, guardrail and approval prototypes
Engineering complexity addressed
Multi-tenant trust boundaries
Organisation context, role enforcement, governed data access, throttling and traceable activity keep sensitive workflows separated and inspectable.
Broader agent-system R&D
Prototypes explore sequenced and parallel specialist agents, cost tracking, PII handling, persisted findings, output checks and failure fallbacks beyond the production file-note core.
My role
Co-founder and technical product lead. I led product architecture, AI workflow design, application development, integrations, infrastructure and the feedback loop between real adviser behaviour and the product.
What this proves
The production file-note core is the strongest proof that I can take AI beyond a demo. The wider R&D shows how I extend that foundation carefully into agents, controls and more ambitious workflows.
Tend
An owner-controlled pet passport and operations network for pet care.
Tend connects an owner-controlled pet health passport with booking readiness and day-to-day operations for kennels, daycare, groomers and other pet-service businesses.
The difficult problem
Pet owners repeatedly rebuild the same health and care record, while businesses receive incomplete information across forms, messages and screenshots. Both sides lose time and businesses still have to judge whether a pet is ready for a specific service.
The product insight
The pet profile should belong to the owner and travel with the animal. Businesses should receive only the information needed for a booking, with provenance, clear readiness rules and sharing the owner can revoke.
What I built
- Consumer web and mobile apps, a provider dashboard, administration and a shared API
- Source-backed document extraction with confidence and human-confirmation states
- Scoped, expiring and revocable sharing between an owner and a business
- Bookings, capacity, waitlists, care tasks, invoices, payments and reminders
Engineering complexity addressed
Portable records without lost control
Random sharing tokens, explicit data scopes, expiry, revocation, ownership checks and source-document hashes let the record travel without becoming public or permanent.
Policy-based booking readiness
Required and recommended health items produce a clear verdict for each service, and active bookings can be re-evaluated when the underlying records change.
A connected facility operating model
Tenants, venues, staff, rooms, capacity, waitlists, Stripe payments, care tasks and reports remain connected as a booking moves through the business.
My role
I drove the product model, consumer and business experiences, shared architecture, booking and readiness logic, payments, permissions, AI records assistant, analytics and the go-to-market operating system around the product.
What this proves
Tend shows that I can turn an everyday frustration into a deep, multi-sided product—connecting trust-sensitive consumer data to business operations while thinking about adoption, payments and commercial constraints.
Stories Alive
A multimodal AI studio built for Australian primary teachers.
Stories Alive turns one teaching topic into a coordinated classroom project—storybooks, songs, animated videos, worksheets, presentations and lesson plans—with a guided assistant helping teachers shape and revise the work.
The difficult problem
Creating a coherent set of classroom media takes far more than one prompt. Language, images, music, voice, video, curriculum context and teacher review all have to agree while expensive jobs fail, restart and compete for resources.
The product insight
AI becomes more useful to a teacher when every resource belongs to the same Teaching Project. Characters, curriculum decisions, language and creative direction can be reused while generation remains visible, resumable and editable.
What I built
- Connected workflows for stories, songs, videos, worksheets, presentations and lessons
- A multimodal pipeline across planning, images, music, voice, lip-sync and assembly
- Lucy, a streaming assistant that can extract structured product actions
- Curriculum provenance, quotas, cost caps and a ComfyUI GPU-worker prototype
Engineering complexity addressed
Character and scene continuity
Role-labelled reference views, film planning and automated clip review protect identity, action, visual style and pacing across many separate model calls.
Resumable, efficient production
Idempotent jobs, partial-progress streaming and a tested low-memory FFmpeg path are designed to keep long projects moving without duplicating expensive work or exhausting a worker.
Curriculum and cost boundaries
Pinned curriculum sources, rights approval, immutable market policies, per-output quotas and account caps constrain what the system can generate before spend begins.
My role
I shaped the teacher product, multimodal generation architecture and creative workflow, then connected the application, background processing, GPU integration prototype, curriculum controls and media review experience.
What this proves
Stories Alive demonstrates that I can coordinate frontier models into a substantial vertical product—with policy, quality, cost and deployment concerns handled around the creative experience.
Agent Layer
Giving AI agents private context without giving the cloud the keys.
Agent Layer is a personal health-data MVP exploring end-to-end encrypted context: a cloud API stores ciphertext and limited routing metadata while a local MCP layer holds the keys and exposes selected context to a user's agents.
The difficult problem
Useful health and wellbeing agents need durable context, but centralising sensitive activity and health records as readable SaaS data creates a new privacy and lock-in problem.
The product insight
The storage provider does not need plaintext. Encryption, key recovery, sync semantics and the agent interface can be designed as one system so context remains useful without surrendering control.
What I built
- A ciphertext API with limited routing metadata and schema-versioned encrypted chunks
- A local MCP daemon that holds keys, decrypts and merges context
- A portable cryptographic core for mobile, web and agent runtimes
- Recovery, relay and sync primitives designed to fail closed
Engineering complexity addressed
Cryptographic key and recovery model
The MVP implements domain-separated keys, authenticated encryption, BIP39 recovery and sealed boxes while keeping independent security review as a launch gate.
Encrypted sync model
Monotonic change feeds, canonical headers, merge-on-write and tombstone rules are designed to preserve consistency for the current MVP without asking the server to inspect contents.
Portable agent access
Shared crypto and schema packages support an Expo app, web console, API and local MCP service without creating incompatible security formats.
My role
I led the product and system architecture across the vault, cryptographic package, sync model, API, mobile and web clients, and the local agent-facing MCP layer.
What this proves
Agent Layer shows that I can work below the application surface—reasoning about cryptography, data ownership, synchronisation and protocol boundaries while recognising that external security review and legal sign-off remain launch gates.
Expat Sidekick
A safer operating desk for life in a foreign language.
Expat Sidekick helps English-speaking expats in Germany understand letters, bills and PDFs, organise related correspondence and prepare editable English and German replies without hiding the source behind a chat response. Forwarded-email workflows remain under active validation.
The difficult problem
Immigration, tax, debt, insurance and health documents combine unfamiliar language, deadlines and legal consequences. A plausible AI summary is dangerous if the user cannot see what was extracted, correct it or control what happens next.
The product insight
Chat can guide the experience, but structured evidence and actions are the real product. Uploaded content stays untrusted data, verification is visible and preparing a reply remains separate from sending it.
What I built
- Photo, multi-page, QR phone and PDF intake, with forwarded email under validation
- Reviewable extraction of dates, amounts, contacts, risks and requested actions
- Issue timelines that group related documents using source evidence
- Bilingual reply preparation, a mailbox flow under validation and official-source monitoring
Engineering complexity addressed
Untrusted documents, bounded actions
Document and email content cannot change system instructions, permissions or memory. Every record stays household-scoped and important facts retain provenance.
Noisy, multi-channel intake
Photos, OCR, attachments and sender names are normalised into currencies, dates, contacts and issues without pretending uncertain extraction is verified truth.
Approval before external action
The implemented flow is designed to separate draft preparation from sending and require a second explicit confirmation with duplicate-send protection; final production QA remains.
My role
I adapted a mature B2B platform into a consumer product, then developed the document workflow, tenant-safe trust model, capture channels, issue timeline, bilingual reply experience and agent boundaries.
What this proves
Expat Sidekick shows how I can reshape substantial underlying infrastructure for a very different audience while keeping evidence, uncertainty and explicit human control at the centre.
Website Factory
A production line for bespoke local-business websites.
Website Factory turns a business URL into a researched, fact-checked and conversion-focused new website, visually reviews the result, publishes it and manages the human-approved outreach around it.
The difficult problem
Scaling creative delivery usually makes the work feel generic. The hard problem was repeatedly producing a distinct, accurate website while coordinating research, code generation, visual review, publishing and careful outreach.
The product insight
AI can accelerate the production line without flattening the output when source facts, design direction, fresh screenshots and human approval remain explicit at every stage.
What I built
- A four-stage workflow from research and grading through design and implementation
- Fact-integrity controls, SEO/GEO foundations and screenshot-led visual QA
- A durable job queue with provider orchestration, retries and restart recovery
- Campaign experiments, guarded outreach and automated AWS publishing
Engineering complexity addressed
Quality without invented claims
The system carries real services, brand assets and source facts into generation, then validates artifacts and uses new screenshots to correct the finished experience.
Resilient multi-model production
Stage-specific models, timeouts, logs, process cleanup, priority lanes and lost-process recovery keep long-running creative jobs inspectable and resumable.
Responsible outreach operations
Inbound deduplication, reply classification, unsubscribe handling, send caps, cooldowns and do-not-contact enforcement sit inside the same product as delivery.
My role
I designed and built the internal product, staged AI workflow, preview and correction loop, durable queue, campaign operations, AWS publishing and the commercial service around it.
What this proves
Website Factory demonstrates that I can productise complex creative delivery—joining AI, design, software operations, quality control, deployment and go-to-market in one live internal system.
The transferable skill
The industry changes. The delivery pattern does not.
I can enter an unfamiliar domain, understand the operating reality, identify the difficult middle and carry a product from idea through implementation and iteration.
Shape the product
Turn a broad problem into a focused proposition, workflow and interface people understand.
Engineer the difficult middle
Handle the integrations, permissions, state, failure paths and data boundaries between idea and reality.
Build for real use
Add observability, review, testing, cost control and operational resilience where the product demands it.
Connect product to market
Think about positioning, onboarding, pricing and the reason somebody will change how they work.
What are you trying to build?
Bring me the complicated version.
A regulated workflow, a new product, an AI system that needs better foundations, or an idea nobody has made concrete yet—I can help shape it and get it working.
Start a conversation on LinkedIn