Build Log

Finance teams at pharmaceutical advertising agencies get surprised by margin-impacting developments — client procurement consolidations, regulatory shifts, competitive pipeline changes, pricing pressure — that were already signaled in publicly available data weeks or months earlier. Internal reporting catches these after they hit the P&L. Nobody was systematically mining public sources to generate predictive, finance-actionable intelligence.

Federated swarm orchestration within Microsoft Copilot, powered by GPT-5.5 (1M-token context). Indication-specific swarms perform specialized due diligence. A central Chairman / Orchestrator synthesizes, challenges, prioritizes, tracks predictions, and feeds closed-loop learning back into the swarms.

Public Data Sources (SEC, FDA, Earnings, ClinicalTrials.gov, Press, M&A)  ↓Source Monitoring Agent → Ingest, classify, log freshness, track cadence  ↓Margin Lever Mapping → Map signals to shared taxonomy  ↓Signal Router (IF) → Revenue at Risk / Pricing Pressure / Out of Scope  ↓Analysis & Prediction  ├── Finance Briefing Agent (standardize for Chairman)  ├── Prediction & Confidence (hypothesis, evidence, time horizon)  └── Evidence DB + Contradictory Evidence Search  ↓Chairman / Orchestrator  Synthesize · Challenge · Prioritize · Cross-indication detection  Route to finance · Manage prediction status · Closed-loop tracking  ↓Finance Outputs  ├── Forecast Review Briefing (monthly, 3 days pre-meeting)  ├── Revenue Risk Watchlist (account-level)  ├── Pricing Pressure Watchlist  └── Prediction Log (status · checkpoint · validation)

• —Human-in-the-loop gate: Every prediction requires human validation before reaching finance workflows. The system advises; it does not act.
• —Closed-loop learning: Predictions tracked through a lifecycle — Open → Confirmed / Partially Confirmed / Disconfirmed / Expired / Superseded — with outcomes fed back into swarm weighting.
• —Anti-obsolescence doctrine: 60-day vendor/platform reassessment cycle. 12-month hard ceiling with conditional 90-day extension requiring 4 named criteria by Month 10.
• —Quantitative decision gates: 5 gates with hard pass/fail thresholds across source sufficiency, automation reduction, cross-indication comparison, governance readiness, and value-to-cost (must exceed 3x project cost).
• —Investment thesis before build: Plan requires identifying 3-5 historical margin events, quantifying dollar impact, estimating detection lead time, and calculating value of earlier detection at conservative mitigation rates.

In MD review (Version 2.0, amended per MD feedback). Pilot scoped to defensive margin protection with reduced 4-agent set. Quantitative thresholds added to every decision gate. Investment thesis and retrospective backtest required before any build activity.

Pitch teams had no structured way to stress-test RFP submissions before presenting to client-side boardroom panels. Reviews were subjective, inconsistent, and biased toward the pitch lead's perspective. The decision-makers had predictable backgrounds, biases, and evaluation priorities. Nobody was modeling for them.

RFP Document  ↓Persona Engine (Myers-Briggs mapped to procurement roster research)  ↓Per-Persona Evaluation (alignment, bias flags, priority weighting)  ↓Composite Score + Gap Analysis  ↓Actionable Debrief (pre-pitch)

Shipped and adopted by pitch leadership. One senior colleague with 20 years in healthcare advertising said he'd never seen anyone think to build something like this.

BLNK is a one-person brand operation. Maintaining voice consistency, content quality, and strategic alignment across channels doesn't scale without systems.

Claude Project (BLNK Brand Ops)  ├── Voice Library Skill (tone, vocabulary, constraints)  ├── Audience Persona Skill (segments, motivations, friction)  ├── Channel Guidelines Skill (IG, email, event comms)  └── Caption Generation Skill (formulas, hooks, CTA patterns)

2 of BLNK's 5 sold-out events occurred after deployment (January 2026). Content velocity increased without sacrificing voice consistency.

New business pitches often required custom product-specific websites to demonstrate speed-to-market and creative capability. Building these sites to align with internal web standards and security practices under pitch timelines — sometimes as short as one week — was functionally impossible through conventional development. Based on prior client services invoices, a custom-coded site of this scope ran upwards of $50K: a full codebase developed, revised, and refined by a human developer. The gap between "we should build a site for this pitch" and "we can actually deliver one" was killing the team's ability to signal bias for action.

RFP / Pitch Brief (ingested)  ↓Agent Core (natural language → site generation)  ↓Custom RAG Layer  ├── UX/UI Best Practices (layout, accessibility, responsive)  ├── Brand Compliance Rules (internal web standards)  └── Security & Deployment Guidelines  ↓Branded Website Output (deployment-ready)

Prototype built and demonstrated. Currently under review by the company's AI Committee for mass deployment. Primary criticism: the tool's interface is too technical for the intended user base — account leadership managing diverse client portfolios with limited bandwidth to learn new agent-based workflows. The feedback pointed to a UX simplification requirement before broader rollout, effectively shifting the deployment bottleneck from "does it work" to "can non-technical users operate it independently."

As BLNK's sole operator — responsible for all content, public communication, campaign management, and media spend — I was also sourcing brand partnerships manually. The research, qualification, and outreach cycle consumed time that should have gone toward building the brand's identity. I needed a system to streamline discovery and eliminate selection bias.

Claude MD Skill Package. Deliberately not an agentic pipeline — architecture chosen to reduce hallucination risk and token cost. Structured skill documents guided partner research criteria, qualification rubrics, and outreach templates.

The system was effective at building shortlists — surfacing aligned partners faster and with less bias than manual research. But end-to-end automation didn't hold. For event companies and creative organizations at BLNK's size, highly personalized communication was the determining factor in whether partnerships closed. The tool couldn't replicate that, and managing the system's continued refinement consumed nearly as much time as it saved.

Pitch cycles operate under extreme time compression. Key stakeholders spent hours on repeatable research and preparation tasks.

Three purpose-built agents, each targeting a specific bottleneck in the pitch preparation cycle:

Q&A Agent  Input: Client brief, RFP requirements, historical Q&A from prior pitches  Output: Pre-drafted answers to anticipated client questions  Integration point: Pitch prep sessions (pre-rehearsal) Medical Due Diligence Agent  Input: Brand/product documentation, regulatory context, therapeutic area data  Output: Medical accuracy checks, compliance flags, citation verification  Integration point: Medical review stage (pre-submission) RFP Insight & Kickoff Agent  Input: Raw RFP document  Output: Structured brief — key requirements, evaluation criteria,          competitive positioning signals, recommended kickoff agenda  Integration point: Day-one RFP receipt (kickoff acceleration)

All three shipped and integrated into active pitch workflows. Reduced hours of manual preparation per cycle across medical, strategy, and account teams. Adopted as part of the founding AI team's initial deployment at Omnicom Health.