Deprex Design Philosophy — When AI Embeds Itself into Your Company

The Forward Deployed Engineer model that Palantir pioneered is neither consulting nor a packaged product. Engineers embed inside client offices, directly touching data pipelines and continuously reflecting "what is actually happening in this operation right now" back into the software. Instead of throwing proposals over the wall from outside, you see problems from inside. That shift creates decisive differences in implementation quality.

Deprex re-implements this FDE model with AI agents. Instead of a human engineer, an AI agent embeds inside the customer's chat tools, email, and document management systems — learning operational context and executing work autonomously. Here's how human FDE compares to AI FDE:

Importantly, AI FDE does not claim to fully replace human FDE. AI has clear advantages in routine work, information collection, and report generation. Legal judgment, building client relationships, and handling genuinely novel situations remain human responsibilities. What Deprex defines is the boundary of "work that can be delegated to AI" — and drawing that boundary is itself the first value delivered.

Andrej Karpathy's question cuts to the core of AI deployment: "Why does AI rediscover the same things from scratch every time?" This isn't a capability problem. It's a memory design problem.

Traditional AI systems rediscover the same information on every query. Customer information organized yesterday, workflow patterns learned last week — all reset to zero in the next session. This "no memory" design keeps AI as a "smart but forgetful assistant" and prevents it from evolving into an "autonomous operator."

Karpathy's answer is straightforward: convert Raw Sources (raw data) into a Compiled Wiki (structured knowledge), and hold it in a Schema (queryable form). Rather than rediscovering on each query, compound the knowledge that was compiled once. This is the foundation of Deprex's memory system design.

Deprex's memory system is built across 3 tiers. Each tier has a clear responsibility, and as data transforms into higher tiers, it crystallizes into reusable, compounding knowledge.

The key design principle is that the Raw Sources tier is immutable. The LLM never rewrites primary records. Only the Compiled Wiki tier is maintained by the LLM — and when new information arrives, multiple related pages are updated simultaneously. This keeps individual information fragments from becoming isolated, instead forming an organic knowledge graph that grows over time.

A Lint cycle maintains consistency: periodic checks detect contradictory entries for the same entity, pages stale for 90+ days that have more recent information available, and orphaned pages with broken cross-references. Knowledge doesn't just accumulate — it stays current and coherent.

However, whether the Raw Sources tier is needed depends on the nature of the primary information source. For static primary sources — books, academic papers, archived documents — the Raw Sources tier works effectively. Since the originals don't change, storing raw data enables future reprocessing and reinterpretation. For dynamic primary sources — Slack messages, API responses, web pages — the Raw Sources tier is likely YAGNI (You Aren't Gonna Need It). Because the originals change in real time, statically storing raw data means staleness begins the moment it's saved, and the agent's exploration capacity is constrained by referencing "old stored raw data" instead of fetching the latest.

The recommended pattern for dynamic sources is to skip raw data storage entirely: fetch on demand via the Connectors tier, let the agent apply semantic judgment to distill important insights, and persist those to the Compiled Wiki. When undistilled information is needed later, fetch the latest from the source again. This preserves exploration capability while still compounding knowledge. The exception is when legal or regulatory requirements mandate raw data retention (tax documents, audit trails, etc.) — in those cases, treat them as static sources and store in the Raw Sources tier.

Deprex's Connectors tier collects data from every tool a customer already uses. No new tooling needs to be adopted. Existing communication, documents, and code management all become inputs to the knowledge system.

The collection policy is fundamentally differential: after the initial sync, only data newer than the last collection timestamp is fetched. This minimizes resource consumption while keeping knowledge current. Collected data lands in the Raw Sources tier immediately, triggering the Ingestion Pipeline to update the Compiled Wiki.

When a business agent executes a task, it runs through a 7-phase cognitive loop modeled on how humans solve problems. This is not simple task queue processing — it's an observe→remember→reason→plan→act→evaluate→learn cycle that improves with each execution.

A key property of this loop is that the EVALUATE phase uses pre-defined Acceptance Criteria as the mechanical judgment basis — not subjective "feels complete." This makes quality assessment possible without human review. What LEARN encodes gets persisted to the Compiled Wiki tier, so the next time a similar task runs, REMEMBER retrieves that knowledge and execution accuracy improves. The loop is self-improving by design.

Deprex is built across five layers. Each layer owns a clear responsibility, and the loose coupling between them means adding a new data source, applying customer-specific rules, or swapping an implementation detail in one layer doesn't ripple across the others.

The practical significance of loose coupling: if a data source changes its API, only the L2 connector for that source needs updating. L3 and above trust only the contract that "normalized data arrives" — they're unaffected by changes in collection method. Similarly, changing customer A's business rules in L5 leaves the shared cognitive engine in L4 untouched. This structural separation keeps per-customer scaling costs from growing linearly with complexity.

The most debated aspect of Deprex's design is the no-approval-gate principle. The 7-phase cognitive loop continues executing tasks without human confirmation at each step. This isn't about trust. It's about design.

Systems that demand constant confirmation stop at every pending approval. The speed at which humans can approve is thousands of times slower than the speed at which AI can process. Without resolving this asymmetry, AI remains an expensive assistant rather than an autonomous operator.

No approval gates doesn't mean every decision is exempt from human oversight. Escalation conditions are pre-defined: irreversible operations, high-impact external actions, and decisions exceeding defined thresholds all escalate to humans — with notification via their preferred channel. "What should humans decide?" is designed upfront, and everything else is delegated to AI. That boundary design is where the real value of Deprex is delivered.

Deployment to a customer environment proceeds in 3 phases. It's not simply installing software — it's AI agents observing, learning, and progressively expanding their scope of autonomous operation.

From Phase 2 onward, a self-improvement loop plays a critical role. Agents discover new automation opportunities during execution, and feed those discoveries back as inputs to the next automation cycle. The goal isn't perfect automation from day one, but learning through use and expanding scope through learning. Knowledge accumulated for each customer is anonymized and stored in a pattern library, making it immediately available for the next deployment.

Deprex's long-term vision is "zero-human operations." This is not something that happens overnight — it's a hypothesis-validation roadmap that advances in stages, with each phase's results serving as the evidence base for the next.

"Zero human" is the north star — a compass, not a destination. What matters is honestly defining, at each phase, what gets automated and what remains human. Phase 0 failure blocks Phase 1. Verifiable milestones built incrementally are more honest than ambitious promises for this class of product. In the early phases, the careful work of distinguishing "what requires human judgment" from "what AI can own" is itself the essential value Deprex delivers.