Jabir | AI Systems

Building production AI agents and LLM systems with TypeScript & Python. Sharing architecture decisions, failures, benchmarks, and code from real projects.

[Jabir | AI Systems @codeCaptain404]

I’m building AI developer tools in public.

Not demos—products with real constraints: local inference, native IPC, developer workflows, resumable pipelines, and human approval.

I share the architecture decisions, tradeoffs, and code behind them.

What I’m building ↓

[Jabir | AI Systems @codeCaptain404]

Automatic GitHub/GitLab CI ingestion is not included yet — that’s a future integration target. For now, you can paste or upload evidence manually.

https://t.co/ZaCeyiclqK

[Jabir | AI Systems @codeCaptain404]

Röntgen AI: Relay build pipeline optimizer

I built Röntgen AI, a build pipeline optimizer for Relay. It accepts uploaded or pasted CI evidence and suggests pipeline improvements.

[Jabir | AI Systems @codeCaptain404]

Repo is public on GitHub.

https://t.co/ZaCeyiclqK

[Jabir | AI Systems @codeCaptain404]

Röntgen AI: Atlas to Blueprint architecture diagrams

I'm working on Röntgen AI: an architecture diagram generator that flows from Atlas to Blueprint. Atlas describes the current system, Blueprint the target. #architecture

[Jabir | AI Systems @codeCaptain404]

The pipeline can take Atlas and help produce Blueprint. You may optionally review intermediate steps. It doesn't replace judgment — it structures the path.

[Jabir | AI Systems @codeCaptain404]

The diff is +158 / -30 across those files. It improves how multi-line comment messages are handled in the prompt flow.

https://t.co/8ydRmvLkcC

[Jabir | AI Systems @codeCaptain404]

IntelliTab: multi-line comment message prompt support

I added multi-line comment message prompt support to IntelliTab. The commit touches python-server/model.py, python-server/server.py, src/completion-provider.ts, and src/context-extractor.ts.

[Jabir | AI Systems @codeCaptain404]

Radar: incident analysis. Relay: CI pipeline optimization. The seven products can be used together or separately; the workflow is a map, not a mandate.

https://t.co/ZaCeyiclqK

[Jabir | AI Systems @codeCaptain404]

Röntgen AI: a seven-product engineering workflow map

I built Röntgen AI as a seven-product engineering workflow map: Blueprint, Pulse, Atlas, Sentinel, Forge, Radar, Relay. Each one targets a different phase of the dev lifecycle.

[Jabir | AI Systems @codeCaptain404]

Blueprint: architecture and security review. Pulse: chat over spreadsheets and SQL. Atlas: repository explanation. Sentinel: PR review. Forge: issue workflows.

[Jabir | AI Systems @codeCaptain404]

Under the hood, the reference architecture uses Appwrite: seventy TablesDB tables, three private storage buckets, and an orchestrator function. This is an optional target, not a constraint. Feedback welcome.

https://t.co/uJrRdNKwDF

[Jabir | AI Systems @codeCaptain404]

OrkestriaAI: durable background execution

I've been building OrkestriaAI, a project for durable background execution. Runs and individual steps are persisted, so failures can be retried instead of lost.

[Jabir | AI Systems @codeCaptain404]

The orchestration model includes idempotency, exponential backoff, timeouts, dead-letter handling, approvals, and audit events. It may look like a lot, but each piece targets a failure mode I've seen in practice.

[Jabir | AI Systems @codeCaptain404]

The target is simple: keep low-risk analysis moving, but make human confirmation non-optional for high-impact changes. Optional automation is fine; blind automation is not.

https://t.co/uJrRdNKwDF

[Jabir | AI Systems @codeCaptain404]

OrkestriaAI: deterministic approval policy

I built OrkestriaAI with a deterministic approval policy. Read-only analysis may run autonomously, but purchases, submissions, deployments, permission changes, destructive actions, and sensitive data transfers pause for human approval.

[Jabir | AI Systems @codeCaptain404]

Approval requirements are computed from action, environment, destination, data classification, cost, and configured policy level. Same inputs, same decision — no guesswork.

[Jabir | AI Systems @codeCaptain404]

https://t.co/ZaCeyiclqK

[Jabir | AI Systems @codeCaptain404]

Röntgen AI: Sentinel and Blueprint security review

Security review assistant is now included in both Sentinel and Blueprint. It's optional and can be run when you need it.

[Jabir | AI Systems @codeCaptain404]

The assistant may help you catch issues before release. You can target specific files or directories, or run it across the whole project.

No extra services needed. It lives in the repo and works with your existing workflow. See the GitHub repo for details.

[Jabir | AI Systems @codeCaptain404]

https://t.co/uJrRdNKwDF

[Jabir | AI Systems @codeCaptain404]

OrkestriaAI: five products, one control plane

I built OrkestriaAI: an AI ops control plane with five products — Vela (browser work), Loom (workflows), Tempo (DevOps), Helio (cloud cost), Aegis (security review).

[Jabir | AI Systems @codeCaptain404]

All five share one control plane for identity, permissions, context, approvals, observability, and audit history. You can adopt one or several.

For me the key is: consistent oversight across different AI tasks, with an audit trail. Early project, but open to thoughts.