CortexBuild — Executive Summary¶
The opportunity¶
Software teams of every size carry two recurring jobs they can't fund with human engineers:
- Modernising legacy code — the 100k-line app that works but rots
- Burning down backlog — the 50 small tickets piling up while seniors handle modernisation
These are the two biggest "we know we need to do it but never get to it" line items in every engineering org. Today they're either: - Ignored (causing slow decay) - Outsourced to expensive consultancies - Handled by hiring junior engineers (slow ramp, high cost)
CortexBuild is a governed builder platform: it renders your team's rules into the AI coding tools you already run and governs the changes they produce on your existing repos, with user-controlled approval gates.
The featured wedges¶
CortexBuild renders your team's rules into the AI coding tools you already use and governs the work they produce on your existing repos. We do not build a new IDE. We do not lock you into one LLM provider. We do not target one industry.
Two surfaces are the featured wedges in v1 — wedges, not walls:
| Surface | Job-to-be-done |
|---|---|
| A. Legacy Modernisation | "Here is a 150k-line app. Modernise it." |
| B. Backlog Burn-Down | "Here are 50 Linear/Jira/GitHub tickets. Work through them." |
Beyond these wedges, CortexBuild also drives net-new work — new features, services, endpoints, and refactors — but always on an existing repo, never greenfield scaffolding from an empty project. Both featured surfaces share one orchestration engine, one rule SSoT pattern, one recurrence ledger, and one approval matrix.
The defensible edge¶
| What competitors have | What CortexBuild has uniquely |
|---|---|
| Single agent (Sweep, Cursor, Cline) | Multi-agent wave model with adversarial Critic |
| Hard-coded prompts | Rule SSoT that renders to Copilot, Claude, Cursor, Codex simultaneously |
| Forget every session | Recurrence ledger — agents learn from past failures |
| Vendor-locked LLM | BYO key — native Claude (Anthropic) · native Gemini · any OpenAI-compatible endpoint (covers OpenAI/Codex models, Bedrock/Vertex via gateway, Ollama). Host tools (Copilot, Claude Code, Codex CLI, Cursor) are render targets, not LLM providers — they run on whatever model their vendor configures. |
| Vendor-locked VM | Daytona-based — self-hostable on your hardware (execution sandbox planned, Phase 1) |
| Industry-specific | Generic enterprise rules that adapt via brownfield analysis |
| Closed-source (Devin) | Open-core strategy (TBD; see 99-OPEN_QUESTIONS.md) |
The math¶
| Metric | Estimate |
|---|---|
| Target customer ACV | $1.2k – $60k/year |
| Cost per modernisation PR | $0.50 – $5.00 LLM tokens (BYO key — user pays) |
| Cost per backlog ticket | $0.05 – $0.80 LLM tokens |
| CortexBuild margin on Cloud Team tier | ~70% (we host UI + control plane; user pays compute + LLM) |
| Time to first PR for a new project | ≤ 30 minutes from "Connect Repo" |
| Time to validated v1 launch | ~30 weeks (4-phase plan in plan/01-PHASED_PLAN.md) |
Risk summary¶
| Risk | Severity | Mitigation |
|---|---|---|
| LLM provider CLI breakage | High | Adapter pattern + daily smoke tests + pinned versions |
| Token cost spiral on large repos | High | Per-agent model tier routing (Planner/Critic → reasoning_heavy; Builder/Tester/Optimizer → balanced; Explore → fast_cheap) + budget gates + graphify-style indexing |
| Bad PR floods erode trust | Critical | Critic gate mandatory + self-audit YAML + confidence threshold |
| Compliance liability if AI does wrong | Medium | User-monitored + ToS shields platform; nothing irreversible without approval |
| Provider lock-in (us to them, them to us) | Medium | OSS engine + multi-provider from Day 1 |
What we need from the user to start¶
Answer the 5 questions in 99-OPEN_QUESTIONS.md (name, repo location, time
budget, first test target, distribution channel). Phase 0 (extraction +
validation) is then 4 focused weeks.