Gatehouse
The problem
Coding agents now write most of the code, and review is the bottleneck: a Harvard working paper across 718 firms found engineers wrote about 30% more code after adopting agents while completed work did not measurably rise and pull requests took 49% longer to merge. In banks, insurers and hospitals the problem is worse — cloud AI reviewers are unusable because the code cannot leave the perimeter, and Goldman Sachs reportedly pulled Claude Code from bankers and traders in September 2026 over data-security and IP-leakage concerns.
The idea
Why now
Three pieces landed this week — SoduWork's acceptance-criteria-plus-independent-review governance model, BRIDGE's minimal agent memory, and DaisyRose-27B, an RL-trained open agent model that doubles its base model's Terminal-Bench score — and the demand side just got measured: a Harvard working paper (Nov 17) shows review is now the bottleneck, and the Goldman pullback (Sept 28) shows regulated orgs are structurally excluded from cloud reviewers.
What it combines
SoduWork (governance: acceptance criteria plus independent AI review plus human sign-off) + BRIDGE (minimal RAG memory that remembers past review verdicts at ~5-7K tokens per query) + DaisyRose-27B (open-weights agent model, RL-trained for long-horizon agent tasks, runnable on-prem) + a typed decision gate for the merge verdict. Each piece alone is infrastructure; together they make a merge decision that is checkable, memorable, auditable, and entirely inside the customer's perimeter.
MVP
Build first: (1) a GitHub check that runs DaisyRose-27B via Ollama against a PR, (2) an acceptance-criteria file the reviewer must check each item against, (3) a BRIDGE-style memory file of past verdicts, (4) a block-until-human-approves rule with an audit log. Deliberately skip: Jira/Linear sync, auto-fix, multi-repo dashboards, hosted anything — one repo, one Docker image, one gate.
How it makes money
The moat
Business model plus architecture. CodeRabbit, Greptile and Qodo sell per-seat cloud review — their revenue requires customer code to flow to their servers. An air-gapped, flat-licence, open-weights gate cannibalizes that model: to copy it they would have to ship models and infrastructure on-prem, support air-gapped updates, and give up per-seat metering — a rebuild of both product and pricing, not a feature. The buyer segment is also structurally excluded from their product today: regulated orgs cannot send source code to a third-party cloud at all.
Distribution
B2B, sold to platform engineering and application-security teams at banks, insurers, healthcare providers and government agencies; the wedge is the air-gapped Docker image they can trial on their own hardware in an afternoon, sold through GitHub Enterprise Server / GitLab self-managed marketplaces and the security resellers that already serve these buyers.
Why it wins
CodeRabbit, Greptile, Qodo and Sourcery all review PRs in their cloud — the code must leave the building. Gatehouse is the only gate that runs the reviewer, the memory and the verdict inside the customer's own perimeter, and it gates on versioned acceptance criteria written before the work starts, not on a summary written after it.
Risks
Biggest risk is reviewer quality: an open 27B model will miss things a frontier cloud reviewer catches. The MVP de-risks it by benchmarking the gate on the team's own historical merged PRs and routing only low-risk paths to auto-verdicts while humans keep high-risk paths.
Build it with
- SoduWorkAcceptance criteria plus independent AI review plus human sign-off — the governance workflow the gate enforces.
- BRIDGE: minimal RAG memory for agentsMinimal RAG memory that remembers past review verdicts at ~5-7K tokens per query, so the same argument is never re-litigated.
- DaisyRose-27BOpen-weights agent model, RL-trained for long-horizon agent tasks, that runs on the customer's own hardware as the reviewer.
- Microsoft Decision 1.1: a GA decision model for high-frequency agent micro-decisionsCheap GA decision model for the merge verdict itself — approve, block, or escalate on every PR.
Repo to start from
gatehouse — self-hosted merge gate: acceptance criteria, an independent open-model reviewer, review memory, and a human sign-off with a full audit trail
Evidence
- When a colleague's AI-generated pull request lands on you
- Harvard paper: AI coding agents shifted the bottleneck to review
- AI code review tools compared (CodeRabbit pricing)
- SonarQube vs SonarCloud: self-hosted pricing
- Goldman Sachs pulls Claude Code from bankers and traders
- Using AI coding tools in a regulated enterprise without exposing your code
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.