Client work

Retinue

Specialist agents for investor outreach, with human control over the send.

Private research implementation

A relationship-led fundraising workflow combining research, drafting, conversation, and intake agents with investor matching, an append-only ledger, and reviewed outbound actions.

Discuss Retinue
Retinue editorial cover with typography and a conceptual software workflow
Generated editorial artwork for Retinue; not an application screenshot.
THE PROJECT

A closer look at Retinue.

Retinue explores how a team of agents can support investor fundraising without treating a generated draft as permission to contact someone. The implementation separates information gathering and proposed content from the outward action. The repository is private and its scenarios and data are synthetic.

The challenge

Investor outreach depends on the history of a relationship as well as the fit of a candidate. A collection of agents can lose that context or bypass review when their tools are composed carelessly. The challenge is to keep matching, drafting, and conversation connected to recorded facts while maintaining a distinct authorization boundary around sending.

The product experience

Four specialist roles divide research, drafting, investor conversation, and founder intake. Matching applies hard eligibility filters before relationship and similarity ranking. Drafts draw on the relationship record, and a review queue preserves escalations. An approval bridge binds a reviewer-issued token to the exact proposed action, consumes it once, and enters the guarded send path.

  • An explicit four-specialist topology with role-specific tool access.
  • A relationship ledger and matching integration that refuse unreadable state.
  • Durable review work and a single-use approval bridge for the guarded send path.

Product & engineering

Python orchestration declares agent roles and tool ceilings through the Claude Agent SDK. Pydantic-based contracts structure content, while an imported Chaperone library supplies the deterministic policy boundary and matching stages. PostgreSQL stores relationship touchpoints with append-only triggers, durable review items, and action-bound approvals. Approval consumption uses an atomic conditional update. The send boundary records confirmed, failed, or unverifiable delivery and escalates missing audit records. Frozen captures and ranking evaluators provide repeatable evaluation; they do not establish production model reliability.

The decisions that shape the work

Separate content from action

Allow specialists to propose useful content while reserving outbound effects for the guarded review path.

Rank from recorded relationships

Apply eligibility before ranking and distinguish an absent relationship from an unreadable store.

Approve one exact action

Bind expiring approval to the recipient, body, tool, investor, and mandate; consume it before execution.

Working within the constraints

  • The implementation and evaluation scenarios use synthetic data; no production fundraising outcomes or autonomous-delivery performance are claimed.
  • The research, drafting, and intake specialists have no outbound send tool; conversation sends are held for human review.
  • Unavailable relationship state must stop the operation rather than silently becoming an empty history.
  • Ladder-derived tiers, contested-quantity rendering, adaptive weight updates, sliding-window contact limits, and store unification remain designed work.

Have a similar challenge?

Retinue illustrates the systems needed around agents that participate in relationship-sensitive workflows: reliable context, specialist coordination, recoverable review, and explicit authorization for external actions. Those patterns apply to other operational software where a draft and a completed action must remain distinct.

Talk about your project
IN THE DETAILS

The work, in view.

Client work

More to explore.

All work
Rezzie editorial cover with typography and a conceptual software workflow
Client workLive product

Rezzie

A guided AI product that turns a trusted resume and a job description into an editable, relevant application draft.

AI SaaSDocument workflowsBrowser editing
Explore the project
ResearchOS editorial cover with typography and a conceptual software workflow
Client workLive product

ResearchOS

A structured research and writing product with inspectable sources, saved workspaces, and cited papers.

Agentic researchSource provenanceCited writing
Explore the project
YOUR NEXT PROJECT

Let's build the software your work needs.

Start a conversation