How we built an AI company brain for ColdIQ.
ColdIQ is an $8M outbound agency running a sales team, a content team, coaching and playbooks, and a membership product its clients log into every day. Every function had its own tools and its own slice of the company's knowledge, and none of them talked to each other. We shipped them one unified platform and put the whole company on it: every campaign, every call recording, every client win, every internal playbook, centralized in a single persistent memory layer that keeps learning as the work happens. A fleet of autonomous agents runs on top. A DRM-locked learning product serves the members. The internal team gets the full knowledge graph. Members get a permissioned slice of it. It runs in production on ColdIQ's own infrastructure, and they own every line of code.
The challenge
ColdIQ does not run on one engine. There is a sales team, a content team, coaching calls and playbooks, and on top of all of it, a membership product their paying clients log into every day.
Every function had its own tooling and its own slice of institutional knowledge. None of it was connected, and the expertise that made ColdIQ worth $8M was fragmented across teams and locked in a few people's heads. Scaling meant hiring more people to carry that context. That was the problem.
One platform, one brain
We built ColdIQ one platform and put the whole company on it. Every campaign, every call recording, every client win, every internal playbook: ingested, embedded, and indexed in a single persistent memory layer that updates as the work happens, because we wired their entire toolstack into one brain.
Under the hood it is a retrieval-augmented generation system: OpenAI embeddings in a pgvector store, hybrid semantic and keyword search, reranked, with Claude reasoning on top. What that means in practice is simple: anyone on the team can ask a question in plain language and get an answer grounded in ColdIQ's actual calls, documents, and frameworks, cited to the source.
An agent layer, two sides
A fleet of tool-using Claude agents runs on that same memory layer. They query eight live data sources, draft content in the founder's voice, and operate Close, Attio, Instantly, and LinkedIn Ads over their APIs from plain conversation.
Because the agents have access to ColdIQ's own playbooks and institutional memory, they do the high-value work too. They build GTM strategies from the campaigns that actually closed, draft proposals in ColdIQ's proven structure, and stand up full outbound sequences, copy included, generated from the internal frameworks that built the business. The expertise that used to live in a few people's heads now runs on demand, at any hour, without a headcount increase.
The same platform serves two sides. ColdIQ's internal team works against the full knowledge graph. Members on the product side get a guided, permissioned slice of it. One platform, two access tiers, and it scales to every new member ColdIQ adds without any additional infrastructure work.
The learning product
The member-facing side is built to teach, not just host content. Every member gets a private workspace, isolated at the query layer so no one's data can surface in anyone else's session. Inside it they build their own knowledge base from their own material, running through the same ingestion and retrieval pipeline as the company's, so the AI answers from their content, not a generic script.
Every lesson becomes a live conversation: ask a question mid-video and get an answer grounded only in that lesson's transcript, with citations rendered as timestamps that seek the player to the exact second. Lessons auto-summarize so members review in minutes instead of rewatching for an hour. Because each person learns against their own workspace, their own material, and their own questions, no two members get the same experience. The personalization is structural, not a setting.
Built not to be copied
Because the course library is the product, we made it genuinely hard to steal. Every video streams through VdoCipher with Widevine and FairPlay DRM, so a screen recording comes out as blank frames, blocked at the hardware compositor layer rather than with browser-level tricks.
A dynamic watermark burns each viewer's own email across the frame, so any leaked copy names the exact account it came from. Playback is authorized one session at a time with a single-use, server-minted token, so the keys never reach the browser and there is nothing in the page to lift. Members watch freely inside the app. The video cannot be downloaded, hotlinked, or leaked without a name attached.
One platform, in production
It all runs in production on Google Cloud Run, with every model call traced and costed to the cent. One custom platform: it runs the agency on the inside, teaches their members on the outside, scales as they grow, and ColdIQ owns all of it.