Technology
Frontier models where reasoning matters. Local inference where privacy matters. Retrieval and provenance at every step — because a finding without a citation isn't a finding.
Architecture
Core Components
Native PDF, scanned image, and DOCX ingestion. OCR for degraded scans; layout detection preserves tables, schedules, and exhibits. Headings, clause boundaries, and defined terms are identified structurally — not inferred from formatting alone.
Semantic chunking at clause granularity. Embeddings from Voyage or local encoders for sensitive corpora. Stored with document, page, and paragraph provenance so every retrieval can be cited.
Hybrid BM25 + dense retrieval surfaces candidate passages; the reasoning model applies the playbook against retrieved context, not against model memory. This is how hallucinations are kept out of the report.
Open-weights models deployed in-VPC or on-prem for sensitive matters. Same pipeline, same output schema — the routing layer decides which inference backend a document hits based on classification and client policy.
Gold-set evaluation harnesses measure extraction precision and recall per clause type. Production monitoring tracks latency, token spend, model drift, and flagged reviewer overrides — feedback loops close automatically.
PostgreSQL for operational data, SQLite for lightweight deployments, S3 for document storage with client-side encryption. Containerized and deployable to AWS, Azure, GCP, or fully air-gapped on-prem Kubernetes.
Why Hybrid
Merger documents, trade secrets, privileged communications — these do not belong in third-party cloud logs. Local inference keeps them on your hardware without sacrificing the analysis.
Bulk classification — "is this an NDA? is this exhibit relevant?" — runs cheaply on local or small cloud models. Frontier models are reserved for the 5% of work that actually needs their reasoning depth.
Different models are better at different things. The architecture routes each task to the model that performs best on it in your evaluation suite — not the one that sounded best in a demo.
I'll review your workflow, constraints, and sample documents, and tell you honestly whether this approach is a fit.
Schedule a Technical Conversation