Deep-dive project documentation — architecture decisions, what broke, how it was fixed, and what shipped. Each study covers the problem, the system design, and measurable outcomes.
Reps advanced deals to contracting and KYC without knowing if the role or industry was even eligible for EOR. Compliance only caught blockers at contract review — after the prospect was ready to sign. Lost deals, wasted KYC, eroded trust.
A 46-node n8n production workflow with three trigger paths: real-time webhook on deal creation/stage change, hourly safety-net sweep, and on-demand batch re-screen. The core architecture decision: the LLM only maps messy CRM text to one of 30 canonical sectors — a deterministic JavaScript matrix makes every tier decision. This means every verdict is repeatable, auditable, and cites its own reasoning.
Edge cases were reasoned about explicitly: pass-through detection (a staffing firm placing workers into oil & gas is screened as oil & gas, not staffing), semantic sector aliasing, and a "needs-title" hold instead of guessing on thin data.
Const-reassignment bug on first live test: caught immediately by the verification step; fixed and re-tested green. Prevention: the dry-run harness now gates every change.
Five analytics properties silently failed to write: didn't exist in the CRM. Folded content into the audit-note field instead. Prevention: validate property existence before writing.
Old prompt let thin data demote clean sectors: re-architected so confidence is a report field that never changes the tier.
Reps spent time after every call on follow-up research and writing. The automation meant to handle it was failing 21 of 25 runs — a dead Azure AI Search index left four AI agents with no country context, producing generic filler instead of specific, grounded output.
Rather than recreate a fragile index (which needed permissions I didn't have), the whole failure point was designed out: extract countries mentioned in the call transcript → ISO-3 codes → fetch each file directly from Azure Blob Storage → inject into every agent. No fuzzy search, no credentials to rotate, graceful fallback so the email still sends if a file is missing.
84% of failures: dead Azure AI Search index (DNS not-found): removed the dependency entirely; replaced with deterministic blob fetch.
Silent dead-ends on scheduled/future meetings: added meeting-completion guard + bounded retry budget (10→20 polls) + explicit exit nodes.
Broken paired-item lineage dropped transcripts downstream: carried the base object forward and converted 28 fragile item-references to a stable form.
Enrichment credit usage spiked from under 15K/month to 515K in a single month. 21% of spend ran through unattributed "Copy of" tables. Phone numbers were bought 2–3x per contact. A single day burned 201,924 credits from a phone-waterfall re-buying the same numbers.
Full-year forensic audit: a seven-sheet master dataset that re-attributed unnamed "Copy of" spend to real source tables. Then shipped a Standard Enrichment Table Template with three build profiles (Lean ~15 / Balanced ~21 / Full ~50 credits per row) and a credit-planning architecture: gate → cap → clone-don't-copy → dedupe → consolidate → weekly review.
Chicago's 48th Ward (Edgewater, Uptown, Andersonville) had no digital way for residents to find their precinct captain, look up local businesses, or submit service requests. The ward office managed 39 precinct contacts in Excel, 1,500+ business licenses as PDFs from the city, and service requests via paper forms. This doesn't scale.
Your48 V1.0 shipped as a static civic homepage with ward directory, city services, community resources, events calendar, and an AI chat widget. V2.0 adds precinct directory with interactive Mapbox map, searchable business license database from Chicago Data Portal, constituent intake forms with Airtable backend, and a public ward dashboard for transparency. The entire platform runs at $0/month on Vercel's free tier.
Multi-agent AI systems suffer from context loss between sessions, naive task routing (keyword matching instead of phase-aware allocation), brittle retrieval (single-pass search without credibility control), and brittle prompting (raw prompts instead of compiled manifests). Production systems need architecture, not ad-hoc chains.
ROSTR formalizes four components: PAL (Prompt Abstraction Layer) compiles loose natural language into strict agent runtime manifests through a five-stage pipeline. RAG DAL (Dynamic Acquisition Layer) performs multi-pass retrieval across three credibility tiers (academic → editorial → community). NPAO classifies work into a 5D phase taxonomy (PreD → Design → Development → Deployment → Debugging) and scores priority on 4 dimensions. The Rostr Hub provides persistent state with four-level management (session → project → org → agent).
The framework is the backbone of the Gency AI agency — turning SOPs into intelligent agents — which cuts idea-to-scoped-project from 4–8 hours to under 30 minutes.
A paid SDR agency needed a steady stream of clean, targeted, compliance-screened prospect lists every week — without contacting existing customers, open deals, closed-lost, or recently-touched contacts. Manual list building was slow, error-prone, and risked burning the agency's time (and the company's reputation) on dead-end or off-limits accounts.
Built a repeatable intake → normalize → dedupe → enrich → suppress → activate pipeline across Clay, Amplemarket, ZoomInfo, and HubSpot. Every company is screened live against the CRM before send: customer status, open deals, closed-lost, and active sequences are all checked. A custom JS round-robin engine assigns each prospect to the right sales pod based on territory (US West/Midwest/Northeast/South-Central + EMEA/Canada + Global).
The pipeline also includes an APAC filter that removes companies from Asia-Pacific markets using TLD detection (`.hk`, `.cn`, `.jp`, etc.), domain keyword matching, and company name analysis — catching APAC records even when HubSpot location fields are empty. Schools, governments, and competitors are filtered out via ICP screening.
Parallel one-off workbooks caused drift: different team members built their own Clay tables. Consolidated into one master workbook with segment tabs as the operating source of truth.
APAC records slipping through with empty location fields: added domain TLD detection, domain keyword matching, and company name analysis as fallback detection layers — catching records that HubSpot location fields missed.
Organizations, SMBs, and consulting agencies have SOPs — documented processes that run their business — but turning those into autonomous AI agents requires deep expertise in agent frameworks, workflow orchestration, and platform configuration. Most teams don't have that in-house, and generic AI tools don't understand their specific business logic.
Gency AI uses a three-step ROSTR-powered process: JTBD Analysis maps the exact job the operator needs done, with deep intake to understand functional requirements. NLP Intent Extraction takes natural language descriptions and maps them to architecture. Then the ROSTR System delivers a full agent, workflow, or skill — production-grade automation the team can adopt instantly.
Service verticals span AI Agent Development (Claude Code, Codex, OpenClaw), Workflow Orchestration (n8n, Make.com), IDE & Platform Setup (Antigravity, Claude Code environments), and Web Apps & Custom MCPs that let agents interact with business logic. The interactive demo on the site lets prospects paste an SOP and watch the ROSTR engine generate architecture in real time.