Six builds across three ideas: agents should run repeatable processes, LLMs should work with meaning rather than templates, and markets should be understood as networks. Each repository has one home.
Current proof
What’s in the repos.
6 shown
01 / Agent-led systems
FinX Intelligence Agent
Private · migrated
A scheduled research pipeline that turns a defined social network into evidence and a weekly market-intelligence product.
More detail
Problem
Chronological feeds reward recency while hiding recurring signals, market narratives and useful disagreement.
What I built
Collection, structured LLM analysis, evidence storage, scheduled execution and report delivery across Python, Supabase and Railway.
Current state
The repository migration is verified. The stronger theme-discovery and follow-up-search loop remains Seb-owned work.
Accounts→Posts→Analysis→Evidence→Brief
Python
LLMs
Supabase
Railway
Private repository · publication review
01 / Agent-led systems
OutboundOS
Private · working slice
A governed operating system for repeatable outbound execution, with explicit operator control before anything becomes action.
More detail
Problem
One-off automations create activity without a durable workflow, shared state or trustworthy approval boundary.
What I built
Scheduled runs, deterministic research fixtures, pending proposals, explicit review, approved materialisation and append-only events.
Evidence
The clean reference slice is restart-safe and covered by six tests. Employer-specific operating logic stays private.
Schedule→Research→Proposal→Review→Output
Python
Approval gate
Event log
6 tests
Private repository · publication review
01 / Agent-led systems
AI Agent Traffic Listener
Private · working slice
A privacy-conscious server-side listener that measures AI-agent visits without creating a new dataset of ordinary human traffic.
More detail
Problem
Browser analytics undercount or obscure crawlers and live retrieval agents, leaving an emerging discovery surface invisible.
What I built
A configurable 29-label registry, unknown-agent discovery, privacy-minimised events, asynchronous writes and an insert-only schema.
Boundary
This measures agent traffic. It does not claim to solve how a company is represented, cited or recommended inside model answers.
Request→Detect→Classify→Event→Report
Node.js
Next.js pattern
Supabase
24 tests
Private repository · publication review
02 / Semantic intelligence
Campaign Resonance Engine
Private · working slice
Find companies whose own language provides evidence that a campaign premise is likely to resonate.
More detail
Problem
Firmographic targeting says who a company is, but rarely whether it is already expressing the problem or worldview behind a campaign.
What I built
Reviewed prospect-language queries, passage-level semantic retrieval, best evidence per signal and breadth-first company ranking.
Evidence
Every result retains its matched passage, label, score and source URL. The safe reference implementation has five tests.
Premise→Language→Search→Evidence→Audience
Python
Embeddings
Evidence
5 tests
Private repository · publication review
02 / Semantic intelligence
Semantic Market Profiler
Private · working slice
Turn an open-ended market question into an operator-reviewed source plan, evidence model and post-qualified result set.
More detail
Problem
Useful market attributes often live in unstructured text and cannot be bought as a clean database field.
What I built
Editable source plans and semantic signals, an approval gate, evidence-backed profiles, post-qualification and natural-language querying.
Portability
The original hiring-philosophy use case is now one configuration; a separate workflow-readiness example proves the wider model.
Question→Source plan→Evidence→Profile→Qualify
Python
Operator control
2 examples
7 tests
Private repository · publication review
03 / Network analysis
GraphGTM
Private · working slice
Map a buyer ecosystem, identify its communities and bridges, then search the people and evidence inside it.
More detail
Problem
Account lists flatten a market that actually spreads through communities, influence paths and connector relationships.
What I built
PageRank, betweenness, Louvain communities, participation and role labels, plus evidence-preserving concept search.
GTM output
The system ranks plausible entry points and explains them through graph position, community role and matching profile evidence.
The working repositories remain private while their final release checks, screenshots and README passes are completed. The portfolio ships first; inspectable public artifacts will follow project by project.