Engineering deep dive
System & moat.
How the system works. Why it gets harder to copy over time.
Scope
Scope.
- Domain: multi-document due diligence
- Input: heterogeneous document sets
- Output: structured, traceable findings
- Goal: reproducible evaluation across datasets
The goal is not summarization — but consistent, reproducible evaluation.
Problem surface
Same fact — told three different ways.
No shared schema. No canonical source. Consistency must be inferred — not assumed.
Input constraints
Input constraints.
- OCR noise (scans, handwritten documents)
- Fuzzy entity names
- Layout-dependent meaning
- Style variability
Inputs are noisy and inconsistent. Extraction is probabilistic — and must be validated.
System objective
Convert unstructured documents into structured findings.
Model produces text. System produces structure.
System model
System model.
Multi-stage pipeline. Each stage produces a structured intermediate state — the input to the next stage.
Every intermediate state is inspectable, testable, replayable.
Execution model
Execution model.
- Stage dependencies
- Schema-bound outputs
- Validation gates
- Retry / fallback paths
Each stage depends on validated outputs from previous stages. Execution is controlled — non-linear when needed.
Control layer
Control layer.
- Deterministic rules (if / then)
- Schema constraints
- Pass / fail logic
- Validation before progression
Rules define what must be true. Models assist — they do not decide correctness.
Replicable components
Replicable components.
- UI / document viewer
- LLM-based extraction
- Basic pipelines
- Simple rule checks
Straightforward to implement. They do not provide defensibility on their own.
Complexity drivers
Complexity drivers.
- Entity normalization across documents
- Cross-document reconciliation
- Second-order validation
- Missing-data inference
- Consistency across datasets
Complexity arises from interactions across documents — and the need for consistent behavior.
Core capabilities
Core system capabilities.
Four primitives. Each feeds the validation core.
Language & jurisdiction
A separate axis — wrapping the pipeline.
Affects meaning — not just data. Where local complexity lives.
Feedback & learning
Feedback & learning structure.
Each transaction extends the explicit knowledge base around the model.
Compounding loop. Not training on data — extending structure around it.
Failure modes
Failure modes.
- Extraction errors
- Entity mismatch
- False conflicts
- Missed inconsistencies
- Missing-data errors
Errors propagate across stages. They must be monitored — and reduced over time.
What becomes hard
What becomes hard.
- Stability across datasets
- Reducing false positives / negatives
- Handling edge cases
- Maintaining behavior under change
The challenge is maintaining consistency — under real-world variability.
Learning mechanism
Learning mechanism.
- Iteration from real cases
- Rule refinement
- Schema updates
- Error correction loops
Learning occurs at the system level — not model training.
Distribution + learning
Distribution + learning.
- Design partners (law firms)
- VC case execution
- Real dataset exposure
- Feedback-driven iteration
Distribution drives system improvement. Each use case contributes to refinement.
Resulting advantage
Resulting advantage.
- Increased reliability
- Accumulated edge cases
- Improved consistency
Advantage builds gradually — through iteration.
Boundary conditions
Boundary conditions.
- Replicable at early stages
- No static moat
- Not protected by model access
The system is not inherently uncopyable. Defensibility develops over time.
System overview
AI reasons. System verifies. Expertise compounds.
Input
- OCR noise / handwritten
- Fuzzy names
- Different styles
- Missing / inconsistent data
System pipeline
Moat primitives → into validation
- Citation lineage (claim → source → region)
- Contradiction detection (mismatch chains)
- Negative assurance (expected ≠ present)
- Section rules (pass / fail)
Output
- Confirmed
- Conflicts
- Missing
- + traceable references
AI reasons. System verifies. Expertise compounds.