Portfolio

Julia
Druck

AI product manager and builder

I'm a senior AI product manager. For the last two years I've been learning generative AI and AI agents, through courses and by building things: SubHeadsUp and ByLevel are both live, and my consultancy runs on a set of agents and automations I specified and built myself, with evals written before anything goes into production. Before that I spent fifteen years in pensions, insurance and energy, where I built the first automated advice product in the UK defined benefit pension advice market.

Worked with

How I build

Spec or prototype, by the cost of being wrong

I match the approach to the risk: a written specification the agents build against where being wrong is expensive, a rough build in front of a real user where it is cheap.

Evals before trust

I write the evaluation set and the guardrails before a feature goes near production, then run them on every change, prompt edits and model swaps included.

Discovery to production

I own the work from discovery and problem framing through design and build to release, and once it is live I keep the evals running, watching for drift and improving it on quality and cost.

AI products and prototypes

SubHeadsUp landing page
B2C product

SubHeadsUp

subheadsup.com

I built and run SubHeadsUp, a subscription tracker that reads renewal signals from Gmail and warns before each charge lands, with no bank login needed. It's live in production, with a Chrome extension alongside the web app.

Full stackAI classificationAgent-built
Aug 2026Open
ByLevel landing page
B2B product

ByLevel

Process mapping with provenance

I'm co-building ByLevel, which draws a map of how a business really works from the documents it already has, tracing every box back to the sentence it came from so anyone can check where a claim originated before acting on it.

ProvenanceDocument AICo-built
Aug 2026Open
AI Governance Co-Pilot
Working prototype

AI Governance Co-Pilot

EU AI Act classifier

I designed and built an agent that tells a small business where it stands under the EU AI Act, prohibited, high-risk or neither, grounding every answer in the regulation's own text with the article cited so the business can act on what it's told.

RAGCitationsRegulated
Feb 2026Read the build
Inside the Agent landing screen
Working prototype

Inside the Agent

Four domains, one failure rig

I built a customer service agent and the rig that breaks it, working across four domains at rising risk, from a garden retailer through GP triage to motor insurance, and measured the guardrails rather than assuming them. Breaking it live showed me that agents hallucinate commitments, not data.

Red teamingGuardrails
Jun 2026

AI operating system

I design and run the AI system Serpin works on.

Runs every day
Running for the business, on a schedule.
Pre-meeting sales brief
Researches the company and the person before a first meeting. Every hard fact is quoted and linked to its source, and my reading of it is kept separate from the facts.
Inbound lead pipeline
Watches for enquiries from three sources, then triages, qualifies and researches each one, files it to the CRM with a verdict on whether it is worth pursuing, and drafts the reply. Nothing sends without me.
Meetings, end to end
Picks up a recording from any of three sources, transcribes it and names who spoke from a voice library I built, then files a written note against the right project, summarised and structured, and flags the ones it is unsure about for me to check. Nothing gets typed up by hand.
The Morning
A notification on my phone each morning pointing at a private page: my diary for the day, what moved yesterday, what is planned today. Composed and published overnight, and nobody else can open it.
AI Radar
Reads ten sources before I am awake, from the labs' own release notes through Hacker News to the business press. It does not hand me a link list: it writes what the day's stories mean taken together, and every item ends with what it changes for a client conversation.
Packaged for other people
Installed into their own Claude in about fifteen minutes.
Mission Control
Five prompts install a self-updating project dashboard into Claude Code and Obsidian.
Sales Recon
Four prompts install an agent team that builds a sourced pre-meeting brief.
Strategy Red Team
Finds the assumption a strategy leans on hardest, then watches for it moving.
Competitor Watchdog
Watches how competitors use AI and flags only what matters.
Personal Sales Video
Scripts and renders a personal avatar video for a named prospect.
Being built now
A multi-agent layer on top, on the Hermes harness. In progress.
A Chief of Staff and a crew
An always-on agent I talk to from my phone, with specialist agents it hands work to.
Work that happens overnight
Scheduled agent jobs that run while I sleep and leave a report I read over coffee.
Receipts and traces
Every agent run writes a receipt, and every step is traced so I can see what it cost and where it went wrong.
Built with
Orchestration
Claude Code · skills · specialist subagents · hooks · scheduled jobs
Retrieval
qmd semantic search · Exa · Firecrawl · Perplexity · NotebookLM
Memory
An Obsidian vault, with a working set and a long-term store kept deliberately separate
Version control
The system and the vault are both git repositories on GitHub, so every change to how it behaves is reviewable
Guardrails
Hooks that refuse destructive commands and block secrets before they can be committed, each with its own regression tests
Observability
Langfuse tracing on my own sessions · a log from every scheduled run
Plumbing
MCP servers · command-line tools of my own · secrets in Infisical, never in code · a private Tailscale network to my own NAS, nothing exposed to the internet

Case Studies

Discovery

An AI opportunity assessment in construction

From site visits to a costed roadmap

I ran a full discovery of a construction business, an industry new to me, with days on site mapping how the work actually flows. I catalogued the opportunities across the business, assessed the strongest in depth and priced a shortlist, each with its return modelled and a phased roadmap for delivery.

DiscoveryROI modellingStakeholders
2026Coming soon
Regulated, at scale · employment

Automating advice on people's pensions

Mercer and Wealth Wizards

I led product on the first automated advice product in the UK defined benefit pension advice market, regulated advice on an approval path that ran through compliance first.

Regulated0 to 1Compliance
2018Coming soon
Enterprise scale

Customer experience design for a new brand

AXA, Moja

I designed and delivered the customer experience for Moja, a challenger brand for a global insurer: the end-to-end journeys, a mobile-first quote and an online account, live in September 2022, across four internal teams and an external agency. Quote conversion up 8%.

EnterpriseInsurance
2022Coming soon

Capabilities

AI product
AI ROI analysis Retrieval grounding Evaluation harnesses Guardrails and reliability Agentic systems in production Multi-agent orchestration Trust and safety design
Product
Discovery and problem framing Roadmap and prioritisation Stakeholder management User research Pricing and commercial cases Regulated and compliance product design Enterprise scale Shipping to production Onboarding and activation Product analytics Customer feedback