Gigabite is a ledger-first transaction platform with AI-native automation at every boundary. AI creates candidate facts; deterministic validators and human review promote them into append-only truth.
The ledger is append-only, versioned, evidence-backed, and auditable.
Model outputs are staged, validated, reconciled, and reviewed before they touch commercial records.
Product attributes, pricing terms, certification claims, payment terms, and shipment facts link back to source documents, pages, spans, tables, or image regions.
Iceberg receives ledger events as a derived mirror. It is never responsible for live commercial truth.
Semantic indexes are derived from documents, candidate facts, and approved records — and can be dropped and rebuilt at any time.
One direction, six stages. Nothing skips validation, and only the ledger commit service writes approved events.
Buyer & supplier uploads, email intake, ERP APIs → immutable object storage.
Classifier → field extractor → normalizer. Every run logged with model, prompt, cost.
Structured facts with confidence scores, each linked to evidence spans.
Schema + business validators, reconciliation, risk & confidence scoring.
Only where required. Queues by domain: specs, matches, certs, credit, disputes.
Append-only events → projections, outbox, lakehouse mirror.
flowchart TB
subgraph SRC[Source channels]
U["Buyer & supplier uploads
CSV · PDF · XLSX · images"]
E["Email intake"]
A["ERP & portal APIs"]
end
subgraph ART[Artifact plane]
S3[("S3 immutable objects")]
REG[("Document registry")]
end
subgraph AI[AI processing]
C["Classifier"] --> X["Field extractor"]
X --> N["Normalizer"]
end
subgraph STG[Staging & controls]
CF[("Candidate facts
+ evidence spans")]
V["Validators"]
RQ["Review queues"]
end
subgraph LED[Operational ledger]
EV[("Append-only events
hash-chained")]
PR[("Current-state projections")]
end
subgraph ANA[Analytics & intelligence]
OB[("Event outbox")] --> ICE[("Iceberg lakehouse")] --> BI["BI · eval datasets"]
end
U --> S3
E --> S3
A --> REG
S3 --> REG
REG --> C
N --> CF
CF --> V
V -->|controls pass| EV
V -->|review required| RQ
RQ -->|approved| EV
EV --> PR
EV --> OB
The write path governs how any fact becomes truth. The price is the fact that matters most commercially — and it commits the same way, three times: a held live quotation, a locked order finalization, and a trued-up delivery settlement. One hardening figure, evidence-linked at every step, never re-guessed.
See the full pricing spine on the Business plan →Ledger events, projections, candidate facts, review queues, credit sub-ledgers, workflow state. Transactional writes, hash chains, row-level security.
SOURCE OF TRUTHOriginal documents, images, parsed pages, model inputs and outputs, evidence files. Immutable.
SOURCE OF TRUTH · RAWBronze → silver → gold: event mirror, entity snapshots, supplier intelligence, training and evaluation datasets.
DERIVEDSemantic search over products, documents, specs, quotes, review history. OpenSearch only if volume outgrows Postgres.
REBUILDABLEOwnership is strict: application services write only through the ledger command API; AI services write only candidate facts, evidence, and model runs; analytical jobs never write back to the operational ledger.
flowchart TB
APP["App & agent services"] --> CMD["Ledger command API"] --> TXN[("Postgres transaction")]
TXN --> EVT[("ledger_event")]
TXN --> PRJ[("projections")]
TXN --> OBX[("event_outbox")]
OBX --> PUB["Outbox publisher"] --> BRZ[("Bronze
raw events & documents")]
BRZ --> SLV[("Silver
entities & snapshots")] --> GLD[("Gold
scores · features · metrics")]
GLD --> DASH["Dashboards"]
GLD --> EVAL["AI evaluation datasets"]
Every domain has an event history and a current-state projection (quote_current, credit_account_current, …). Projections update idempotently, in the same transaction as the commit where consistency requires it.
Customer SKUs, spec versions, required attributes, the SKU equivalency graph.
Factory capability, certification scope and expiry, audit history.
Versioned quotes with itemized cost components, assumptions, validity windows.
Customer and supplier POs, container milestones: booked, departed, cleared, delivered.
Net terms, credit accounts and draws, receivables, double-entry postings, exposure.
Quality issues, dispute records, compliance exceptions, supplier risk history.
Every service has exactly one reason to exist and one thing it may write. The ledger command service is the only writer of approved events — everything else stages, reviews, projects, or publishes.
identity & tenant
document service
review service
notification service
AI processing
candidate facts
retrieval (pgvector)
review triage
ledger command ★
projection service
quote · order · finance
compliance service
lakehouse publisher
outbox processor
eval dataset builder
Every event hashes its canonical payload together with the previous event hash for its aggregate. Tampering anywhere breaks the chain everywhere after it.
erDiagram
SOURCE_DOCUMENT ||--o{ DOCUMENT_PAGE : contains
SOURCE_DOCUMENT ||--o{ EVIDENCE_SPAN : supports
DOCUMENT_PAGE ||--o{ EVIDENCE_SPAN : locates
EVIDENCE_SPAN ||--o{ CANDIDATE_FACT : proves
CANDIDATE_FACT ||--o{ REVIEW_TASK : creates
REVIEW_TASK ||--o{ REVIEW_DECISION : resolves
CANDIDATE_FACT ||--o{ LEDGER_EVENT : promotes_to
LEDGER_EVENT ||--o{ EVENT_OUTBOX : publishes
LEDGER_EVENT ||--o{ PROJECTION : builds
Every candidate fact runs the same gauntlet. The only variable is whether a human sits between scoring and commit — and that is decided by policy, per fact type.
flowchart TB
D["Document ingestion"] --> CL["Classifier"] --> EX["Field extractor"] --> NM["Normalizer"] --> CF["Candidate fact writer"]
CF --> VA["Validators"] --> RC["Risk & confidence scorer"] --> RT{"Review
required?"}
RT -->|no| CR["Commit request"]
RT -->|yes| HR["Human review task"]
HR -->|approved / edited| CR
CR --> LC["Ledger commit service"]
EX -.-> MR[("Model run log")]
NM -.-> MR
HR -.-> FB[("Reviewed examples")]
FB -.-> MR
| LAYER | DIRECTION | NOTE |
|---|---|---|
| Web app | Next.js · React · TypeScript | Operator workbench + buyer portal |
| Backend | FastAPI · Pydantic · SQLAlchemy | Typed contracts end to end |
| Workflows | Temporal | Durable multi-step commercial workflows |
| Operational DB | PostgreSQL + pgvector | Ledger, projections, retrieval in one plane at MVP |
| Objects | AWS S3 | Immutable artifact store |
| Lakehouse | Iceberg · Glue · Athena | Parquet tables; dbt + Spark transforms |
| Events | Postgres outbox → Redpanda | Stream only when throughput demands it |
| Runtime | ECS Fargate · Terraform | Containerized, IaC from day one |
| Observability | OpenTelemetry · Metabase | Traces + model run logs; MVP dashboards |