WHO I AM

Hi, I’m Rob.

12 years in venture capital, strategy consulting, and startup operations. I spend my days investing in AI companies and my nights building agents.

Things I'm Building

Things I Write About

Connect on LinkedIn

The world has changed:

AI changed the rules of the game. The new world demands a step change in agent development and orchestration skill, that only comes from rolling up your sleeves and building. This site is where I share what I've built, and what I'm thinking and writing about as we navigate this new era.

Things I'm Building

Three agents, three different architectures and levels of complexity

Production-grade diligence automation — an 8-pillar Digital Maturity Assessment

Digital Maturity Assessment Synthesis Agent (PE / M&A)

Problem

Diligence teams conducting a digital maturity assessment on a target company can spend weeks on synthesis: reading stakeholder interviews, mapping evidence to a rubric, scoring and sizing opportunities, preparing an executive report. The DMA synthesis agent accelerates and standardizes this process, evaluating targets across 8 pillars (e.g. Data Foundation, Tech Architecture, Process Automation).

Approach

An 8-node LangGraph pipeline takes an engagement package (i.e. structured C-suite interview notes, data profiling tables, app integration inventory) and produces a complete readout: an evidence-grounded score across all 8 pillars, targeted follow-up questions for any pillar where the evidence is too thin, and a sized, thesis-weighted list of the top value-creation opportunities.

Result

Produces 8 output files per run - full report, executive summary, scored heatmap, risk register, initiative backlog, pillar narratives, and process brief - in under 15 minutes.

Architecture

LangGraph, 8 nodes: Intake → Evidence Synthesiser → Pillar Scorer → Follow-up Generator -> Follow-up Integrator → Benefits Sizer → Output Generator → Report Assembler
Self-directed follow-up loop generates role-specific follow-up questions to address evidence gaps
LangSmith/Langfuse tracing on every run
Full eval suite confirms scorer consistency and rubric alignment against known-answer bundles

Video Walkthrough

Architecture diagram

Download Select Outputs

Multi-agent ‘research-to-publish’ system for thought leadership content

LinkedIn PoV Agent

Problem

Producing a well-researched LinkedIn article historically meant stitching together deep research, multiple cycles of critique and iteration, adjustments for tone matching the article intent / target audience - a manual workflow reliant on multiple tools and hours.

Approach

Six coordinating agents on a shared LangGraph state: an Orchestrator decomposes the topic into research elements, a Researcher and Critic run an iterative loop per element (the Critic scores depth, accuracy, and actionability, and routes weak research back for another pass), a human-in-the-loop checkpoint lets the user review research intent before it begins; a Writer and Editor integrate the diverse research components into a final piece.

Result

A streamlined article production agent. 8-12 minutes per run, producing seven output files including the final article, all research documents, and a process brief documenting every decision the pipeline made. Can either be published as is, or used as a '90% done' base, that users can tweak in their preferred LLM before publication.

Architecture

LangGraph StateGraph with a cyclic conditional edge (Researcher ⇄ Critic) - live web research per element, not vector/RAG retrieval
Human-in-the-loop interrupt - user reviews and edits research intent before the loop starts
Six agents, one shared PipelineState - each with clearly scoped read/write responsibility
Structured JSON at every hand-off; graceful degradation on rate limits

Video Walkthrough

Architecture diagram

Download Sample Outputs

Investment-grade research - theme in, ranked shortlist out

Thematic Stock Screener

Problem

Screening public markets for stocks that fit a specific thematic lane (e.g. "Physical AI / robotics") means manually pulling a candidate list and financials, iterating on your analysis and building a prioritized view - a few hours of grunt work before the actual decision-making can start.

Approach

A five-step pipeline: the user gives a theme and optional filters (market cap, region); the agent researches the space via live web search, pulls up-to-date financial data per candidate, scores and ranks every company, and writes a structured research report.

Result

Produces a ranked JSON shortlist with per-dimension scores plus a human-readable markdown report, in minutes. Tested end-to-end on multiple themes (e.g. Physical AI/robotics, Sovereign AI Infrastructure).

Architecture

Python + Anthropic SDK (tool use / function calling) + yfinance + Pydantic.
A five-step linear pipeline with three distinct tool calls: web research, financial data fetch, scoring
Every tool output schema-validated - no free-text parsing
Full cost + latency logging across the pipeline

Video Walkthrough