Collective OS·AI CODING AGENTSWORKFLOW DESIGNHUMAN-IN-THE-LOOPPM × IMPLEMENTATION

AI-Agent Workflows for Relationship Intelligence.

Product implementation · AI-assisted delivery · Sanitized case study

This documents hands-on product and implementation work on Collective OS AI workflows — including workflow design, requirements definition, implementation review, test-case generation with AI coding agent support, and explaining implementation decisions to business stakeholders. It is not a claim of sole authorship of any feature.

Sanitized case study. The workflow UI, data, names, scenarios, and examples throughout this page are synthetic, and implementation details are generalized.

This is a sanitized reconstruction of real product work. The workflow UI, data, names, scenarios, and examples are synthetic, and implementation details are generalized. Product strategy and scope decisions are on the full case study page.
The Business Problem

Relationship data was fragmented and disconnected from action.

Professional networks contain high-value signal — shared contacts, prior interactions, company context, meeting history. The problem is that this signal is scattered across email threads, calendar data, CRM records, and connection metadata, with no layer that connects it to what a user actually needs right now.

  • Users couldn't identify which relationships were relevant to a specific goal without manual research
  • Even when the right person was found, users lacked context on why they mattered or how warm the connection was
  • Choosing an appropriate next action — and explaining it to others — required re-assembling fragmented signals manually
  • AI-assisted recommendations without explanation or human review created trust and accuracy problems
My Contribution

What I built, reviewed, and owned.

The following areas reflect my direct involvement. Language is intentionally precise — I have used hedged wording where I contributed alongside others.

Contributed to

Partner Dashboard

Contributed to the lifecycle of the Partner Dashboard — from initial requirements and user flows through acceptance-criteria review, edge-case mapping, and stakeholder walkthroughs.

Worked on

BYON (Bring Your Own Network) MVP

Worked on the BYON MVP, defining the import-and-enrichment flow, the minimum set of relationship signals to ingest, and the guardrails needed before exposing match results to users.

Helped implement

Matching, scoring, and agent-related workflows

Helped implement and review the candidate-ranking logic, scoring parameters, and the evidence-explanation layer that surfaces reasoning to human reviewers before any action is taken.

Owned

Code review, testing, and stakeholder communication

Reviewed generated code against product requirements, tested failure modes and permission paths, generated test cases with AI coding agent support, and translated implementation decisions into plain language for business stakeholders.

AI Coding-Agent Workflow

How AI agents supported delivery.

AI coding agents accelerated implementation exploration and test-case generation. They did not replace requirements, review, or judgment.

  1. 01

    Convert business problem to structured requirements

    Translate stakeholder goals into user flows, acceptance criteria, states, and explicit edge cases before any code is written.

  2. 02

    Use AI agents to accelerate implementation and testing

    Leverage agents to draft implementation options, generate test-case outlines (integration states, permission paths, error states), and unblock technical questions — then review each output against the ACs.

  3. 03

    Explain implementation to business stakeholders

    Translate technical decisions and trade-offs into plain language — what was built, what was excluded, what the system does when things go wrong, and what success looks like.

AI assistance supported development. Prompts, repositories, and production architecture are not described on this page.

How the agents were built

Five signals in, one explainable recommendation out.

A sanitized overview. Implementation details are generalized.

01 · Signal sources

Five sources feed the platform.

  • Gmail API
  • Recall.ai
  • Exa
  • Apify / Sales Navigator
  • Postmark

02 · Claim extraction

A Claude Haiku claim-extraction pipeline.

Signals pass through a claim-extraction pipeline running on Claude Haiku, so every recommendation can point back to the evidence behind it.

03 · Heat Score

A 0–100 score built from three components.

  • Recency
  • Frequency
  • Depth

04 · Reconnect Agent

Four chained skills, two human approval gates.

The Reconnect Agent chains four skills to recommend who to reconnect with. Nothing moves forward without a person approving it at the shortlist review and again at the outreach draft review.

Shipped

BYON (“Bring Your Own Network”) went live. The platform network is approximately ~1,000 agencies. No outcome figures are claimed on this page.

Product groundwork

A 17-competitor × 14-dimension benchmark and a six-epic roadmap shaped scope before the agents were built. The detail is in the full case study.

Sanitized Demonstration

The workflow — step by step.

An interactive walkthrough of the relationship-intelligence workflow using synthetic data. Click through each step to see the system in action.

Sanitized demonstration · Synthetic data only · No real user information

  1. →
  2. →
  3. →
  4. →
  5. →
  6. →

Step 1 — Opportunity / need

User defines the goal

GoalFind the right person to intro us to Series A FinTech investors in the GCC region
ContextLooking for warm introductions — ideally via mutual connection with existing relationship
Required relationship strengthMeaningful prior interaction (meeting, project, or referral)

The workflow starts with a structured goal — not a free-text search. This shapes what signals the system considers relevant and what evidence it needs to surface before passing to human review.

Step 1 of 7
Testing & Review

What I tested and how.

Three testing areas applied across the AI workflow implementation.

  • Requirements and workflow state coverage

    Verify implementation satisfies defined ACs. Test every state: empty, loading, partial data, error, and success — not just the happy path.

  • Human approval and override behaviour

    Confirm no action is taken without explicit approval. Validate that flagged, dismissed, and overridden decisions are logged and the system does not retry dismissed items.

  • Edge cases and explanation quality

    Test missing data, conflicting signals, and near-zero-confidence candidates. Review whether surfaced evidence actually supports the score — flag mismatches.

Status

Bring Your Own Network went live.

The workflow was developed for Collective OS product and operational use. No adoption or outcome figures are claimed here.

Measurement

What success was designed to measure.

These are the measurement definitions used during design — not claimed historical outcomes.

  • Time to useful shortlist

    Median time from workflow start to a shortlist the user acts on — not just views.

  • Recommendation acceptance rate

    Proportion of shortlisted candidates where the user initiated outreach or logged a next action.

  • Shortlist-to-introduction rate

    Proportion of approved recommendations that resulted in a warm introduction or confirmed next-step action.

No quantitative outcomes are claimed on this page.