SentrAI Roadmap

From the working hackathon MVP (27 Sep 2026) to a product a Singapore SME or private school can run in production.

← Home · Full plan · MVP code

How to read this: 9 phases, each with a goal, the features it adds, a done-when check, the owner, and a rough effort in developer-days for one developer (Erick). Effort figures are estimates made from the current codebase, not measurements; treat them as ±50%. Phases 2 to 6 can overlap once Phase 1 is done.

Where we are today (MVP, done)

Timeline at a glance

Sep   Oct               Nov               Dec          Jan–Feb 2027
P0 ■
      P1 ■■■■
          P2 ■■■■■■■■
            P3 ■■■■■
                        P4 ■■■■■■■
                        P5 ■■■
                            P6 ■■■■
                                          P7 ■■■■■
                                                       P8 ■■■■■■■■

Critical path: P1 (persistence + auth) → P2 (real collectors) → P4 (real containment) → P8 (pilot). Everything else can slip without blocking the pilot.

Phase 0 · Finish the hackathon (28–29 Sep)

Goal: submit a strong video. No new product features.

Done when: video submitted; deck and repo links on the submission.

Phase 1 · Harden the core (≈10 days)

Goal: SentrAI survives restarts, knows who is using it, and every change is tested automatically.

Done when: killing and restarting core mid-incident keeps the incident, penalty and active blocks; an unauthenticated request to any write endpoint gets 401; CI is green on main.

Depends on: nothing. Unblocks: everything.

Phase 2 · Real data collection on every layer (≈18 days)

Goal: monitor real servers, not just the demo portal (plan section 4).

Done when: a Linux VM, a Windows VM, a Postgres DB and an AWS test account all report into one SentrAI; each layer’s example attack from plan section 5 opens the right incident.

Depends on: 1.1, 1.2. Later: rewrite the collector in Go or Rust as a single binary (plan section 13), ≈8 extra days, only after the Python version is stable.

Phase 3 · AI classification that can be trusted (≈10 days)

Goal: prove Jev’s accuracy before letting it influence autonomous actions (plan section 6).

Done when: a published accuracy table exists; false-positive rate on benign traffic is under an agreed target (proposed: under 1% of benign events open an incident); autonomous actions only use rules or Jev results above the calibrated threshold.

Depends on: 2.x for real data (the eval set can start in parallel).

Phase 4 · Real containment and tripwires (≈16 days)

Goal: replace “recorded (simulated)” actions with real, reversible enforcement, and add the tripwires (plan sections 8 and 11). Every adapter follows the same contract: dry-run → snapshot → apply with TTL → verify → rollback.

Done when: on the test VMs, each playbook applies, verifies, expires on its TTL and rolls back cleanly; the Computer Misuse Act rule holds: no action ever touches a system we do not own.

Depends on: 2.1–2.3 (agents on hosts), 1.2 (auth for agent commands).

Phase 5 · Notifications and escalation (≈7 days)

Goal: the right person hears about it on a channel they actually read (plan section 9).

Done when: an unacknowledged incident reaches the operator, then the team lead, then management, on at least two channels, and every delivery and acknowledgement is in the audit chain.

Depends on: 1.2, 1.4.

Phase 6 · Accountability and evidence reporting (≈9 days)

Goal: reports that protect the company and treat staff fairly (plan section 10), and that answer the PDPC’s question: did the organization make reasonable security arrangements?

Done when: a sample monthly board report and a sample incident evidence report can be handed to a PDPC-style reviewer without explanation; signature verification detects any altered record.

Depends on: 1.1, 1.4, 5.x.

Phase 7 · Agents as separate services (≈12 days)

Goal: turn the agent design (plan section 12) from classes in one process into independent services, so a compromise of one agent cannot approve its own actions.

Done when: stopping Countersign stops all autonomous actions (fail-safe), and no single agent key can both propose and approve.

Depends on: 1.x, 4.1.

Phase 8 · Product, pilot and security of SentrAI itself (≈21 days)

Goal: a first real customer, and confidence that SentrAI is not itself the weakest link.

Done when: the pilot runs for 30 days with agreed accuracy and response-time targets, and the pentest has no open high-severity findings.

Differentiators and where they are built

Risks and how we handle them

Team owners

Last updated: 28 Sep 2026.