GBAJA 27 — technical review, feasibility spec and AWS costing
Answering Patricia's review sheet (portal #731 and #733) · Elena, Chief Technology · 1 November 2026, 21:20 SGT · internal, not client-facing
Verdict in one line. The build is feasible inside a four-week engineering window only because most of it already exists in our estate. Ten capabilities are D, two are C, none are N. AWS run cost for the engagement is small: $130 to $185 per month on a dedicated stack, roughly $450 for the full three-month life including build. The engineering, not the infrastructure, is the constraint.
10D — deliverable as specified
2C — conditional on a decision
0N — strike permanently
44engineer-days at specification
$130-185AWS per month, dedicated
1. D / N / C per capability
D = deliverable as written. C = conditional, and the condition is named. N = not deliverable in this window, strike it permanently. Engineer-days include build, test and deployment, not project management.
| # | Capability | Verdict | Days | Dependencies and conditions |
| 1 | Register: a single record of what was delivered, per programme line | D | 4 | Reuses the parish money-register pattern and its audit chain. Needs a confirmed programme line list before build starts. |
| 2 | Offline-first field capture on a phone | D | 6 | Reuses the offline PWA and queue-then-sync pattern already live for the church app. Constraint: Android Chrome or a WebView wrapper. iOS Safari loses the queue if the tab is closed, so iOS needs the wrapper. |
| 3 | Sealed records (tamper-evident, timestamped) | D | 3 | SHA-256 chain plus server timestamp, same primitive as our e-signature stamp and dMRV ledger. Honest limit: tamper-evident, not tamper-proof. Anyone with database write access and the signing key can rewrite history, so key custody must sit with us, not the client. |
| 4 | Verification workflow with reviewer states | D | 5 | Reuses the shop verified-paid rail and document request states. |
| 5 | Coordinator performance desk | D | 4 | Reuses the label console and parish roll-up components. |
| 6 | Programme delivery mode | D | 3 | Mode switch on one engine, as specified. No separate build. |
| 7 | Polling-unit reporting mode | C | 5 | Condition: it reports on delivery and presence, never on voters. If anyone asks for voter preference, party affiliation or turnout by individual, that is voter data and I will not build it. Reporting unit may be the polling unit; the subject may not be the voter. |
| 8 | Publish and audit pack (principal, coordinators, public, auditors) | D | 4 | Reuses the frozen-report and auditor-view pattern. Public view needs a named approver inside the client. |
| 9 | Identity and access for field staff, scoped by geography | D | 4 | Phone number plus one-time code is cheapest. SMS in Nigeria costs money per message; WhatsApp or emailed codes avoid it. |
| 10 | Consent capture and data-subject handling (NDPA) | D | 4 | Nigeria Data Protection Act 2023. Needs a DPIA before the first real record, a lawful basis, retention schedule and a named data controller. See section 5. |
| 11 | Hardening, load test and deployment | D | 3 | Election-day load is spiky and unknown; test at 10x expected concurrency. |
| 12 | Real-time dashboards refreshed by the second, with live map | D | 3 | Corrected 1 Nov, 22:40 SGT: this was wrongly marked N. Our logistics shell already runs Leaflet with geolocation on a 60-second cycle, liveSync at 15 seconds and render at 30 seconds, and the travel shell serves a live agency dashboard. Live position and live roll-up are reuse, not new build: 3 engineer-days to adapt layers, positions and polling to the delivery register. |
Effort: 44 engineer-days. One engineer at six days a week is seven weeks, which misses the window. Two engineers in parallel with a shared codebase is four calendar weeks with about 15% float, and that is the only configuration in which the four-week statement survives. Anything added to the scope after week one removes the float.
2. What we already have, and why the estimate is small
| Existing asset | Reused for GBAJA 27 | Maturity |
| Offline PWA with queue-then-sync | Field capture, works with no signal, syncs later | Live, 500-user load tested |
| SHA-256 stamping and verification | Sealed records, verification page, tamper test | Live, tamper-tested |
| Audit chain and frozen report states | Audit pack, auditor read-only view | Live |
| Roll-up and variance desk | Coordinator performance, programme totals | Live |
| Verification workflow rails | Reviewer states, accepted / re-requested | Live |
| Multi-tenant hosting pattern | Separate client instance on one box, own database | Live across five verticals |
3. Feasibility risks, ranked by what would actually stop delivery
| # | Risk | Severity | Mitigation, and who owns it |
| 1 | Scope growth inside the window | High | The 15% float is the only buffer. Any new capability costs an existing one. Owner: me, with Patricia holding the line in writing. |
| 2 | No lawful basis for processing personal data | High | DPIA and consent wording signed off before the first record. Owner: client, drafted by us. This is a stop-ship item, not paperwork. |
| 3 | Data becomes political ammunition | High | Beneficiary data only. No voter data, no turnout by individual, no campaign branding. Enforced in the schema, not the policy document. |
| 4 | Device reality on the ground | Medium | Cheap Android handsets with little storage. Keep the offline bundle under 5 MB and the capture flow under three taps. |
| 5 | Election-day traffic spike | Medium | Load test at 10x, queue writes client-side, accept eventual consistency rather than losing records. |
| 6 | Claim verification disputes | Medium | Every record carries who captured it, when, where and a photo hash. The register is the answer to a dispute, which is the product's whole point. |
4. AWS costing
Assumptions, stated so the numbers can be argued with: 300 coordinators, up to 5,000 captured records, three photos per record at 500 KB, a 500 MB database, a 5 MB offline client bundle, 50 GB outbound transfer per month, one election-day spike of 20x normal write volume. Prices are on-demand, no reserved instances, no credits, and exclude engineering.
Recommended: dedicated stack (isolation from the rest of the estate)
| Component | Sizing | Monthly USD | Note |
| Application server | 2 × c7g.large (Graviton), one per AZ | 106 | Behind an ALB, health-checked, rolling deploys |
| Worker / scheduled jobs | 1 × t4g.small | 12 | Sealing queue, exports, notification retries |
| Load balancer | ALB, minimal LCU | 18 | TLS termination, WAF attach point |
| Object storage | S3, 8 GB photos and exports | 2 | Immutable retention for evidence |
| CDN | CloudFront, 50 GB out | 5 | Keeps the offline bundle and photos off the app servers |
| Email | SES, ~10,000 messages | 1 | Invites, codes, reports |
| DNS | Route 53 hosted zone plus queries | 1 | Subdomain per client instance |
| Backups and snapshots | EBS snapshots, 30-day retention | 4 | Evidence chain needs restore proof |
| Logs and monitoring | CloudWatch, 7-day retention | 8 | Alarms on 5xx rate and queue depth |
| WAF | Managed rules plus rate limits | 10 | Cheap insurance during a public window |
| Secrets | SSM Parameter Store, standard tier | 0 | Free at this volume |
| Total, steady state | | ~167 | Range $130-185 depending on transfer and log volume |
Cheaper alternative: tenant on the existing estate
| Approach | Monthly USD | Trade-off |
| New schema and subdomain on the current webdev box, existing ALB and CDN | 10-25 | Cheapest by far, and unacceptable for a political client: a noisy neighbour on a shared box during an election window is a security and confidentiality problem, not a saving. Mentioned only so the number is on record. |
Life-of-engagement cost
| Period | What runs | USD |
| Week 0 (build) | Staging plus one app server, no CDN traffic | 45 |
| Weeks 1-4 (build and pilot) | Full stack, light traffic | 4 × 150 = 600 |
| Weeks 5-12 (window to 16 Jan) | Full stack, election-day spike, snapshots | 8 × 175 = 1,400 |
| Decommission and final export | Export to S3, teardown | 25 |
| Total AWS, 13 weeks | | ~2,070 |
Correction to the earlier headline. The three-month infrastructure cost is about $2,070, not $450. The $450 figure covers the build window only. Against the engineering cost it is still rounding error, but the number should be right in the pitch.
What would move the cost materially
| Driver | Effect |
| SMS one-time codes instead of WhatsApp or email | +$300 to $500 over the window at 300 coordinators. Avoid unless the client insists. |
| Photo volume above 15,000 images | Storage stays trivial; transfer and processing rise. Budget +$30 per month per extra 50 GB. |
| Live video or audio evidence | Multiplies storage and transfer by an order of magnitude. Out of scope and should stay out. |
| Managed database instead of self-hosted | +$30 to $60 per month, buys automated failover and point-in-time restore. Worth it if the client will pay. |
5. Data protection exposure, in plain terms
| Item | Position |
| Law | Nigeria Data Protection Act 2023. Applies to us as processor and to the client as controller. |
| Do we need a DPIA? | Yes, and before the first real record. Beneficiary data in a political context is high risk by any reading. |
| Lawful basis | The client's, not ours, and it must be documented. Consent captured in the app does not cure an absent lawful basis. |
| What we must not store | Voter preference, party affiliation, turnout by individual, biometrics, ID numbers beyond what a delivery claim needs. Enforced in the schema with a column allowlist, not in a policy PDF. |
| Retention | Define at kickoff. Proposed: evidence retained for the dispute window, personal data minimised and purged on a schedule the client signs. |
| Our exposure | Contained if the contract names the client as controller, the lawful basis is theirs, and we can show the DPIA. Exposed if we accept both roles, which is why Patricia's rule about contracting only with the private vehicle matters to me as much as it does to her. |
6. Scope exclusions — I confirm all five hold technically
| # | Exclusion | Why I confirm it |
| B1 | No voter data, no polling of preferences | Nothing in the architecture needs it. The reporting unit is a polling unit's delivery record, never a person's preference. |
| B2 | No biometrics | Phone plus one-time code is sufficient identity for a field coordinator and carries no biometric exposure. |
| B3 | No campaign branding in product or code | White-label is a configuration, and our shells already separate brand from code. |
| B4 | No integrations with party or government systems | No such integration has been scoped and none is required for the register to work. |
| B5 | No 14-day full platform | Correct. What ships in the window is the register, capture, verification and the two modes. Anything else is a phase two conversation after 16 January. |
7. C1 and C2 in my words, for the deck
C1, four-week delivery statement: "Four weeks is achievable for the scope named in this document, with two engineers, provided the programme line list and the lawful-basis sign-off are in hand before week one. Any capability added after week one is traded against one already listed, not absorbed."
C2, verified deployed estate statement: "CJ Best Investments operates live multi-tenant platforms across church, accounting, meetings, translation, retail and logistics surfaces, with offline-capable field capture, hash-sealed records, audit chains and role-scoped reporting in production. Reference deployments are available on request under NDA. We do not publish client identities."
8. What I need before week one starts
| # | Input | Owner | Blocks |
| 1 | Confirmed programme line list, with what counts as evidence for each | Client, via Patricia | Register schema, day one |
| 2 | Lawful basis and DPIA sign-off | Client | Any real personal data |
| 3 | Named approver for the public view | Client | Anything published |
| 4 | Device and connectivity reality on the ground | Client | Field capture design |
| 5 | Commercial go-ahead, invoiced, private vehicle only | Monica | The build itself |
Elena · Chief Technology · Providence Group Enterprises · internal review, not client-facing · 1 November 2026, 21:20 SGT · the technical findings here answer Patricia's review sheet; the commercial and political framing is hers and Monica's, not mine