Zack GregoryIndependent technical project · Signal North
NorthTrace
Powered by Veyra AI. An independent exploration of how governed data and AI-assisted investigation can explain differences between enterprise reports.
Interactive lab · synthetic data
Trace the revenue gap
Two systems report the same period in USD. Inspect the source rows, test the explanation and decide whether the evidence supports closing the case.
Reference design
Where an agent fits
A production design would let a model propose an explanation while SQL establishes the amounts. Access controls, evidence identifiers and review decisions travel with the case.
01Enterprise sources Provenance
Capture the source, business key, reporting period and extraction time. Align populations before comparing totals.
02Governed data platform Access control
Apply identity-aware access to curated evidence tables. Preserve source identifiers and lineage without forwarding unrestricted source data.
03Quality and semantic layers Definition checks
Test source completeness and compute measures under an explicit reporting contract. Record the measure version and calculation window.
04Evidence retrieval Scoped context
Select authorized exception rows, relevant metric definitions and execution metadata. Each evidence item carries a stable citation.
05Agent orchestration Bounded investigation
Choose from approved read-only checks and record tool calls. Stop on missing evidence, exceeded budgets or unauthorized requests.
06LLM reasoning Evaluation
Draft an explanation from retrieved evidence. Evaluate grounding, unsupported attribution and the ability to abstain against labeled cases.
07Validation and review Human oversight
Recalculate totals deterministically. A named reviewer examines the evidence and approves or rejects the proposed disposition.
08Business insights Resolution record
Store the decision, owner, evidence identifiers and residual variance. Keep pipeline latency, failed checks and model runs observable.
Current experiment
In my local prototype, DuckDB isolated one $300,000 synthetic invoice missing from BI. Qwen3 4B through LM Studio summarized the SQL results. The upstream cause remained unknown. This browser exercise uses a separate, two-exception dataset so visitors can test the controls.
Research directions
Agent orchestration, local model inference and evaluation are areas of ongoing exploration. The next technical question is whether a model can consistently cite the right evidence and abstain when a cause is ambiguous.