Engineering roadmap · May–Nov 2026 · four-lane view
What engineering will be doing for the next six months
Four lanes — the canonical domain model we are designing, the new commercial platform that implements it, the cloud surface where it runs, and the Hamburg on-premise tier underneath the V3 biological model. Each lane is read left-to-right across closing Q2, Q3, and the Q4 entry.
Colour = lane. Badge = risk band of the work item. Highlighted chips are the highest-leverage commitments. Dashed chips are gated — only landing if their upstream gate clears.
Closing Q2 · in flight
May – Jul 2026
Base layer must land. Notebook is the unlock.
Q3 · differentiation
Aug – Oct 2026
Productized science. Companion Diag. MVP window.
Entering Q4 · outlook
Nov 2026 →
Conditional scope. Decisions due during Q3.
Lane 1 · Conceptual
Designed domain model
V2 bounded contexts & invariants. packages/domain — conceptual only, no DB or UI shape.
Lock the seams
Finish bounded-context split off monolithic projectdesign
Engagement / ServiceSession / AssayOrder lockclear
SampleLineage parentage resolvedopen
InvestigationFrame frameType enum frozenopen
Hard invariants ratified (req ≠ exec, evidence ≠ interp.)clear
Versioned outputs & evidence
EvidenceSet / DerivedEvidenceSet semanticsopen
BenchmarkCorpus / CorpusVersion / CorpusPolicyopen
FindingSetVersion ↔ ReportDeliverableVersion seamclear
CostEvent append-only modelclear
Forbidden-parentage tests in CIclear
Deferred / open
Farm / portfolio grouping — revisitdefer
Interpretation ↔ Decision-automation seamdefer
Recommendation aggregate shapedefer
Trial-intelligence Mode B value objectsopen
Lane 3 · Online tier
Cloud infrastructure
Small / fast / cheap derived layer. Neon · Vercel · GCP · Inngest · Cloud Workflows.
Stabilise & consolidate
Neon prod-v2 · per-PR branch DBs · PgBouncerlow
query-server promoted to sole DB / PostGIS accesslow
Export server greenfield (blob + magic links + DB lifecycle)mid
Geodata server greenfield (raster / vector / TiTiler)mid
GCP Batch · Snakemake hardenedlow
Split exec backbone; tighten observability
Cloud Workflows for sci/ML runs · Inngest for app-coupledmid
Mapbox → Maplibre migration in packages/mapsmid
Sentry / observability package consolidationlow
Cost monitoring · GCP (€397/mo) + Vercel ($127/mo) guardrailslow
Cloud Run stats/ML hardening (AWC / CEC / SOC)low
Forward-looking / optional
Temporal evaluation — kept openeval
SOC 2 evidence pack from sub-processorslow
Latency mitigation on Neon cold-start pathslow
Multi-region readiness checkdefer
Lane 4 · Raw biological tier
On-premise infrastructure
Hamburg petabyte-scale storage + EU GPU cluster. Substrate for V3 generative biology track.
Build the box. Use the clock.
EU GPU cluster · 5,000h · use-it-or-lose-it window opensrisky
Hamburg petabyte storage · racks + network + ingestmid
FASTQ archive lifecycle (GCS → on-prem cold tier)mid
Raw-vs-derived layer split contractmid
V3 training campaign
V3 soil-genomic foundation training (Tim R.)risky
Reproducibility / checkpointing / lineagemid
Data egress / ingress between Hamburg ↔ GCPmid
Bus-factor mitigation in V3 stackmid
Operate & govern
V3 model evaluation + claim provenancerisky
Storage governance · retention · access policymid
Hand-off path V3 → V2 commercial surfacedefer
DR / restore drill on petabyte tiermid
Highest-leverage commitment
Gated · only lands if upstream clears
Low risk · clear / easy
Mid risk · diffuse / medium
Risky · very unclear / very hard
Very high risk
Design decision
Deferred / forward-looking
Central enabling bet
AI leverage on diffuse / medium-difficulty work is the precondition for delivering this scope at current headcount. If the throughput gain doesn't materialise, scope reduces — pace doesn't change.
Q2 → Q3 gate
Scientific customisation must productize in Q2. If Internal Notebook + reasoning capability don't land, Companion Diagnostics slips into 2027 — not Q4. Owner: Maurice + Julia.
Pilot decision · End of Q1 latest
Named Ag-input pilot must be picked before Q3 build starts. Compressing Companion Diagnostics into Q2 = bespoke R&D, not platform. Owner: Bruno + Julien.
Headcount & bus factor
4 engineers; Haydn join-or-no decision pending. Some R&D domains at bus factor 1 — revisit at Series A. Platform-core ownership handover before EOY.