Compiled Awareness (Karpathy Wiki)
A PostToolUse hook rebuilds a ~3k-token index.md on every edit, so every session starts with full situational context instead of blind grepping.
Exadel Code Fest 2026
Slop-Masters
Three solutions. One thesis: AI is only as good as the structure you give it.
The Team
How We Collaborated
We worked as a single team across all three solutions — no per-product owners. We split daily into pair-and-rotate sessions, met morning and evening to compare notes, and kept one shared thesis pinned: AI is only as good as the structure you give it. Each solution is a different altitude on that idea.
Anyone could (and did) commit to any of the three.
Solution 1 of 3
ClearGate
Bridging vibe coding and enterprise delivery.
ClearGate is the disciplined orchestration framework that turns Claude Code from a chaotic junior developer into an enterprise-ready engineering team. It is a planning framework, not a build tool — it scaffolds how AI agents plan and ship work, then hands execution back to your existing CI/test/deploy stack.
-
-
Three Ambiguity Gates
Proposal → Epic → Story. Every phase requires human ratification (approved: true) before the AI can proceed.
-
Four-Agent Execution Loop
Architect (one milestone plan) → Developer (one story per commit) → QA (independent verification) → Reporter (sprint retrospective). Sub-agents never speak directly.
-
Cryptographic Audit Trails
Every Proposal, Epic, and Story is tied to a Git SHA — exact attribution of AI work, with explicit version control on every artifact.
-
MCP Sync to Linear / Jira / GitHub Projects
Local .cleargate/delivery/pending-sync/ is the source of truth; once approved, the MCP adapter pushes ratified Epics and Stories into your existing PM stack.
Solution 2 of 3
Tee-Mo
Secure, Slack-native collaborative AI for the modern workspace.
Tee-Mo replaces the single-player browser-tab bottleneck with a multiplayer workspace where AI lives directly inside Slack channels and the entire team shares, refines, and acts on the same context in real time.
-
Read
Deep context from Google Drive and local files via auto-generated, self-healing file catalogs — no stale caches, no shredded RAG chunks.
-
Act
External execution via MCP — query GitHub, draft Jira tickets, build synthesis pages directly in the workspace wiki (Karpathy Parity).
-
Automate
Scheduled cron tasks deliver background intelligence straight to Slack, with rich execution history and safe dry-run previews.
-
Semantic Routing, not RAG
The "Librarian" model: read AI-generated catalogs of your 15 most critical files, select the right document, and read it entirely fresh. Perfect accuracy, zero staleness.
-
Bring Your Own Key (BYOK)
You hold the API keys; Tee-Mo is purely the engine. Every credential is encrypted at rest with AES-256-GCM. Zero in-process state.
-
Channel-Level Isolation
Each channel gets its own dedicated brain wired to its own documents and tools. Engineering's bot physically cannot reach Marketing's files.
Solution 3 of 3
Exa
An AI-native workspace companion that lives inside Slack.
Exa is the production-ready Slack-native AI brain we built for Exadel itself. Engineers @mention Exa in any channel and ask questions grounded in the team's own documents, PRs, and pipelines.
-
Zero-Setup Ingestion
Drop documents into Google Drive — Exa parses, indexes, and syncs PDFs, Word docs, spreadsheets, and presentations. No engineering required.
-
Project-Isolated Knowledge
Every project gets its own secure data boundary. Team A cannot see Team B's docs.
-
PR & Pipeline Intelligence
Real-time Slack alerts for PR reviews, approvals, and merge conflicts. Smart batching of CI/CD failures cuts noise without burying real breakage.
-
Daily & Weekly Digests
Morning summaries of merged PRs, open reviews, pipeline status, and document updates. Leadership gets clean velocity reports without status-meeting overhead.
-
LLM-Wiki Knowledge Architecture
Documents aren't a messy pile of text — they're parsed into summaries, concepts, entities, and cross-references, structured for LLM-native retrieval.
-
Modular, Cloud-Native Microservices
Slack bot, sync engine, LLM backend, question workers, batch processors — each a standalone service. Runs on a single VM (~$13/month on GCP) or scales to Kubernetes.
The Common Thread
Three altitudes, one philosophy: structure beats magic.
| Axis | ClearGate | Tee-Mo | Exa |
|---|---|---|---|
| Layer | Agent orchestration | Workspace AI fabric | Org-wide knowledge brain |
| Surface | CLI + MCP | Slack + MCP | Slack + MCP |
| Knowledge model | Karpathy Wiki (index.md) | Karpathy Parity synthesis | LLM-Wiki structured graph |
| Trust model | Human-gated, Git-attributed | BYOK + zero-state | Per-project isolation |
| Status | Planning framework, ready to use | Sovereign engine, BYOK | Production-ready |