Core Concept & Value Proposition

FoundryBrain is a SaaS platform that acts as a persistent, AI-powered process memory for metal casting foundries. It captures casting run parameters, defect observations, and shift notes in plain language, then uses a retrieval-augmented AI to correlate historical process data with defect outcomes — surfacing root cause hypotheses in seconds instead of days. Think of it as a shared brain for the entire foundry floor: one that never retires, never forgets a shift, and gets smarter with every pour.

The core promise: reduce scrap rates by 15–30% and preserve the institutional knowledge of your veteran metallurgists before they walk out the door.

The Problem

Foundries bleed money through casting defects (porosity, shrinkage, misruns) that can take days to diagnose. Root cause analysis is slow because:

• Process knowledge lives in the heads of aging, retiring workers
• Shift handoffs lose critical context ("we bumped the pour temperature at 3am because humidity spiked")
• Process logs are scattered across paper sheets, Excel files, and whiteboards
• There is no searchable history connecting "what we did" to "what went wrong"

MVP (Version 1 — Buildable in 8–12 Weeks)

Core features of the MVP:

• Run Log Form — A simple web/mobile form where operators log each casting run: alloy, mold ID, furnace temp, pour temp, binder ratio, humidity, cycle time, and any observations. Voice-to-text input supported for floor use.
• Defect Tagging — When a defect is found, the operator tags the run with defect type (porosity, cold shut, shrinkage, etc.) and severity. A photo upload is optional.
• AI Defect Query — Engineer types "We're seeing macro-porosity on part #A204 — what's causing it?" The AI searches the foundry's own run history using RAG (retrieval-augmented generation), finds correlated parameter patterns, and returns a ranked list of likely causes with supporting evidence from past runs.
• Shift Handoff Summary — At end of shift, AI auto-generates a plain-English handoff brief summarizing anomalies, parameter deviations, and open issues for the incoming crew.
• Knowledge Capture Interview Mode — A structured AI-guided Q&A that helps veteran metallurgists narrate their process knowledge, which gets stored as searchable, citable institutional memory.

Tech stack (lean):

• Next.js frontend, Supabase (Postgres + pgvector for embeddings), OpenAI/Anthropic API for LLM reasoning, Vercel deployment. Two developers can build this.

How AI Is Used

• RAG over proprietary run history — Every log entry is embedded and stored in a vector database. When a defect question is asked, the AI retrieves semantically relevant historical runs and reasons across them.
• Shared persistent context — Inspired by emerging multi-agent context platforms, FoundryBrain maintains a rolling shared context window per facility — so every shift, every user, and every query builds on a unified process memory rather than isolated sessions.
• Anomaly correlation — Traditional ML (simple statistical correlation + LLM-powered interpretation) flags when a parameter combination has historically preceded a specific defect type.
• Natural language interface — No dashboards to configure. Engineers and operators interact in plain English, lowering the adoption barrier on the factory floor.

Unique Selling Points

• Niche-specific, not generic — Purpose-built for foundries with defect taxonomies, metallurgy terminology, and casting process structures baked in — not a generic AI chatbot bolted onto a spreadsheet.
• Gets smarter with your own data — The AI's value compounds over time as the foundry logs more runs; it becomes a competitive moat the customer owns.
• Solves the retiring expert problem — Foundry owners have a visceral fear of losing 30-year veterans. FoundryBrain frames itself as "capturing their knowledge permanently" — a high-urgency, emotionally resonant value prop.
• No hardware required — Pure SaaS, no sensors or IoT integration needed for MVP. Operators log manually. This eliminates the #1 barrier to adoption in manufacturing tech.
• Quantifiable ROI — Scrap rate reduction is directly measurable in dollars. A foundry doing $10M/year with 5% scrap saves $150K–$450K annually with even a modest improvement — making a $500–$1,500/month SaaS fee an obvious decision.

Business Model

• Monthly SaaS subscription per facility: $500 (small foundry, <50 employees) to $1,500 (mid-size, 50–300 employees)
• Onboarding fee ($2,000–$5,000) for knowledge capture sessions and historical data import
• Target customer: Independent foundries in North America and Europe — there are ~1,800 in the US alone, mostly family-owned, underserved by enterprise software

Potential Impact

FoundryBrain turns the foundry floor's most underutilized asset — decades of process experience — into a queryable, AI-powered competitive advantage. For an industry facing a workforce crisis and margin pressure, this is not a nice-to-have; it is a survival tool.

Core Concept & Value Proposition

AdOpilot is an AI-powered co-pilot for ad operations teams that layers on top of existing ad servers — starting with Google Ad Manager (GAM) — to eliminate the manual, repetitive work of campaign trafficking, pacing monitoring, and optimization. Instead of replacing the ad server, AdOpilot becomes the intelligent interface in front of it: operators describe what they need in plain English, and AdOpilot translates that into correctly configured line items, targeting rules, and pacing adjustments — and proactively flags problems before revenue is lost.

The core pitch: "Your ad ops team moves 10x faster without hiring more people."

The Problem It Solves

Ad operations is one of the most underappreciated bottlenecks in digital publishing. Key pain points:

• Manual trafficking is brutally slow — setting up a single campaign with multiple line items, creative sizes, geo targets, and frequency caps can take 30–90 minutes per campaign
• Pacing issues are discovered too late — underdelivery or overspend is often caught manually hours after it starts costing money or goodwill
• Optimization requires scarce expertise — knowing when to adjust floor prices, prioritize line items, or tweak targeting is a specialist skill most small/mid publishers can't afford to keep in-house
• Reporting is fragmented and manual — pulling performance data across line items to write a weekly report is a copy-paste slog
• Small and mid-tier publishers are underserved — they run on lean teams (sometimes 1 person) but face the same trafficking complexity as large publishers

MVP Feature Set

The MVP is a focused web app with a GAM API integration and an LLM-powered chat/action interface:

• Natural Language Line Item Creator — user types "Set up a 728x90 leaderboard campaign for Acme Corp, $15 CPM, targeting US + Canada, running June 1–30, 500k impressions" and AdOpilot generates the correctly structured line item draft in GAM for one-click confirmation
• Pacing Alert Dashboard — connects to GAM reporting API, polls campaign delivery data every hour, and surfaces any line items that are significantly under or over-pacing with plain-English explanations and suggested fixes
• AI Trafficking Checklist — before a line item goes live, AdOpilot runs an automated QA check (correct creative sizes mapped, click URL present, impression tracker valid, targeting not conflicting) and shows a pass/fail summary
• Plain-English Campaign Summary Generator — one-click generation of a weekly performance report narrative from GAM data, ready to paste into a client email

How AI Is Used

• LLM (Claude or GPT-4o) parses natural language campaign briefs and maps them to GAM's API data model (order structure, line item types, targeting keys/values)
• Rule-based + statistical ML monitors pacing curves against expected delivery trajectories and triggers alerts with severity scoring
• LLM for reporting synthesizes raw GAM metrics into readable client-facing summaries with anomaly callouts
• The AI always shows its work — every suggested action includes a short explanation and requires a human click to execute, building trust incrementally

Unique Selling Points

• Zero switching cost — plugs into GAM via OAuth, no migration, no new ad server, no disruption to existing workflows
• Built for lean ad ops teams and solo operators — not an enterprise platform, priced and designed for publishers with 1–5 person ops teams or freelance ad ops consultants
• Plain English in, correct GAM config out — removes the "you need to know GAM's data model deeply" barrier
• Proactive, not just reactive — pacing alerts before campaigns miss goals, not after
• Fast to build — GAM API is well-documented, LLM integration is straightforward; a team of 2 engineers can ship the MVP in 6–8 weeks

Go-To-Market

• Target users: Independent ad ops consultants, small/mid digital publishers (1M–50M monthly pageviews), newsletter operators running their own direct-sold ads
• Distribution: Product Hunt launch, AdMonsters/AdExchanger community forums, LinkedIn outreach to "Ad Operations Manager" titles, partnerships with ad ops consultancy networks
• Pricing: $99–$299/month per GAM network, tiered by number of active line items monitored — simple and predictable for small teams

Potential Challenges & Mitigations

• GAM API rate limits — MVP starts with read-heavy features (monitoring, reporting) and batches write operations carefully; prompt users to confirm bulk actions
• AI accuracy on trafficking — always show structured preview of what will be created before executing; include an "undo" log; accuracy improves with user feedback loop
• Enterprise competition — avoid competing with Placements.io or large AdTech suites; stay laser-focused on the underserved SMB publisher segment and price accordingly
• GAM API deprecation risk — diversify to Kevel and Broadstreet as second/third ad server integrations post-MVP to reduce single-platform dependency