The know-how your company runs on, captured before it walks out the door.
Sparse preserves the decisions only your best people know how to make.
Awaiting query
Type a customs question. Sparse drafts a cited answer for broker review.
Ask a follow-up, inspect a source, or request broker sign-off...
Human signs. AI drafts. Review before filing under licensed-broker supervision.
The risk
The risk isn't losing the documents.
It's losing the judgment that knew what to do with them.
When the one person who held a call in their head moves on, the knowledge leaves with them — but the obligation to act with reasonable care does not. 19 U.S.C. §1484 “reasonable care” outlives the expert who knew how to meet it.
Incoming materials
Source
Critical decisions live across email, docs, and memory.
A thread holds the context, but not the reusable judgment.
Chat
Source
New hires keep asking the same judgment calls.
The answer repeats because the reasoning was never captured.
Ruling memo
Source
Only one person knows why we price this way.
Exceptions and rulings sit with a single license holder.
Spreadsheet
Source
The call that mattered is not one place you can audit.
Signals are visible; the deciding logic is scattered.
Judgment brief
Preserve why the firm knew what to do with the documents.
Obligation
19 U.S.C. §1484
Reasonable care outlives the expert who knew how to meet it.
Supervision
19 U.S.C. §1641
Responsible control still needs a licensed human accountable for the call.
Boundary
19 CFR 111
Sparse keeps the judgment in-house instead of sending it to a public LLM.
Control point
ReviewHuman review remains the control point; Sparse captures the reasoning trail around it.
The reframe
Don't buy “AI agents.” Buy the judgment — before it leaves.
We capture how your best people decide, cite every answer to a source, gate the risky calls behind a human signature, and compound that knowledge inside your firm alone. The model is one component. The harness that makes it defensible is the product.
One governed judgment layer
Cited to your private rulings, gated behind a human signature, compounding inside your boundary. Scoped to customs brokerage — not a generic assistant.
Proven on public data — before we ask for a single record of yours.
ATLAS (arXiv 2509.18400) tests models on 10-digit HTS classification — the exact statutory work a senior broker does by judgment. A fine-tuned ~70B beats frontier reasoning, at ~1/5 the cost, self-hostable.
Fully-correct 10-digit HTS classification
ATLAS benchmark · score is percent correct (higher is better)
Sparse
LLaMA-3.3-70B · fine-tuned
GPT-5-Thinking
frontier reasoning
Gemini 2.5 Pro
Thinking
Scores measure exact 10-digit classification on public benchmark data; results are directional and vary by use case.
Best result
Fully-correct HTS
40%
fine-tuned ~70B
Cost profile
Compared to frontier
~1/5
fraction of cost
Deployment
Control boundary
Self
hostable
Authority
19 CFR 111
Human
reviews + signs
Benchmarks include
What this proves — and what it doesn’t.
Proves (directionally)
A right-sized, fine-tuned model can beat frontier reasoning on a focused task at a fraction of cost — self-hostable, on public data, demo-able without a customer.
Doesn’t prove
Deployment-grade accuracy. 40% fully-correct = a reviewable, cited first draft on every case and a correct first draft on ~40%. It does not replace the licensed broker, who reviews and signs every entry (19 CFR 111). ATLAS is a single self-built study — cited as directional, not an industry standard.
Lower cost, self-hostable, on public data — single self-built study; directional, not industry-standard.
Search retrieves. A chatbot answers. Neither carries your liability.
A category view, not a brand fight. File search finds documents; a chatbot answers from public data. Only a governed brain cites your private rulings, gates the risky calls, and keeps an audit trail that belongs to your firm.
Where it fits
Category view · not a brand fight1. User question
A request comes in.
File search
Included
Covered in this step
Chatbot
Included
Covered in this step
Sparse
Included
Covered in this step
2. Retrieval
Find relevant information.
File search
Included
Covered in this step
Chatbot
Included
Covered in this step
Sparse
Included
Covered in this step
3. Answer generation
Create an answer from sources.
File search
Not included
Stops before this step
Chatbot
Included
Covered in this step
Sparse
Included
Covered in this step
4. Citations
Attach sources and explain the logic.
File search
Not included
Stops before this step
Chatbot
Not included
Stops before this step
Sparse
Included
Covered in this step
5. Policy gate
Check risk, privilege, and rules.
File search
Not included
Stops before this step
Chatbot
Not included
Stops before this step
Sparse
Conditional
Risk checked and gated
6. Human review
Broker reviews and approves.
File search
Not included
Stops before this step
Chatbot
Not included
Stops before this step
Sparse
Conditional
Reviewed and approved
7. Audit trail
Record, store, and replay the decision.
File search
Not included
Stops before this step
Chatbot
Not included
Stops before this step
Sparse
Included
Recorded with full audit trail
Four pillars — the exact requirements of the problem, not a feature list.
The knowledge was never written down because it’s competitive advantage. Governance isn’t overhead — it’s the regulatory requirement. Each pillar maps to a reason the incumbents fail.
System architecture · top to bottom
Sparse framework
Data layer
Explicit docs + the judgment traces in email, chat, and your systems of record — synthesized into one entity-resolved brain.
Governance runtime
focalConstraints, context management, permissions, approval gates, audit replay. Governance is the entry ticket, not overhead.
Provenance layer
Every answer cites which email, which case, which ruling — with an authority weight. An uncited answer is a liability, not a feature.
Vertical model
A fine-tuned ~70B beats frontier on focused task families at ~1/50 params, ~1/100 cost — self-hostable, no egress.
Human review
A licensed broker reviews and signs every entry before filing. The brain drafts; it does not replace the license (19 CFR 111).
Cited output
A cited first draft and a defensible audit record — the only thing that leaves the boundary. Reconstructable end-to-end.
See the thing
A governed, cited brain — made visible.
Real product surface, built on the spec — high-contrast, theme-aware, no images.
Request
Classify: carbon-steel self-tapping screws, zinc-plated, 3.5 mm, boxed 1,000.
Proposed HTS
7318.15.8000Thread-forming (not bolts); zinc-plated finish routes under 7318.15 per NY ruling Nnnnnnn. No ADD/CVD flags on this lane. Boxed-quantity spot-check noted for Brownsville.
Evidence packet · 4 cited sources
“Screws of iron or steel, thread-forming, zinc-plated, not threaded over their full length, are classified under 7318.15…”
view provenance →
Confidence
0.87
Authority
0.91
Risk
Amber 3Low–Med
Senior broker must sign before filing
19 CFR 111.4 · reasonable care · licensed-broker supervision
Human signs. AI drafts. Every claim cited to a source you control.
Your customs records stay inside your boundary. Sparse brings the governed brain to them.
Sparse maps broker inboxes, SOPs, prior entries, and public authority into a permissioned knowledge layer, then returns cited draft work for licensed review.
- 1Before: scattered & risky
- 2Map, retrieve & verify
- 3After: cited & defensible
Before: scattered & risky
Inboxes, SOPs, entries, and rulings are useful — but disconnected, uncited, and off the audit record.
Client documents
invoices, BOMs
Broker SOPs
local playbooks
Team notes
edge-case calls
Broker inbox
classification Qs
CROSS rulings
public authority
After: cited & defensible
A permissioned knowledge layer links every draft to its source trail — inside your boundary, under review, on the audit record.
CROSS rulings
public authority
HTS notes
chapter logic
ABI exports
entry history
Broker SOPs
local playbooks
Proposed lane: 7318.15.8000 for zinc-plated self-tapping screws.
Evidence aligns across public authority, prior entries, and firm notes. Route to a senior broker before filing.
Sources (3)
Human signs. AI drafts.
Boundary-first
Designed for read-only access and in-boundary deployment so regulated records stay controlled.
Permission-aware retrieval
Answers follow existing access rules, with source scope visible to reviewers.
Cited draft work
Every claim carries source trail, authority type, and review state for broker inspection.
Human sign-off
Sparse prepares the reasoning packet. Licensed staff review and sign before filing.
Your data stays inside your boundary. We bring the brain to it.
Governance is the product, not an add-on. The same guarantees scale from a managed SME tier to your own VPC.
Boundary-first architecture
Sources are read where they already live. Sparse returns cited, reviewable outputs without sending private records to shared infrastructure.
Your data & systems
Documents & files
Cloud storage, shared drives, SOPs
Databases & warehouses
Snowflake, BigQuery, Postgres
Sparse reasoning layer
Parse, retrieve, and gate work inside your tenant.
Understand
Parse, chunk, and embed inside your environment.
Retrieve
Find context with hybrid search and reranking.
Review gate
High-risk work waits for licensed sign-off.
Your boundary
Outputs cross only with citations, logs, and review state.
Cited outputs
Answers
Defensible and ready to review
Reports
Exportable with full citations
API responses
Programmatic and traceable
Audit trail
Query logs and source trace
Your data stays in your environment
Single-tenant or customer-VPC. Isolation at the data layer, not just app code. Your private correspondence and the senior's inbox never leave your control — no egress to shared infrastructure.
Answers cite sources
Every quantitative or factual answer links back to its source document and passage. No unsourced superlatives. The reader can verify the chain before acting on it.
Permissions follow existing access
Default-deny, role-scoped. The brain can only surface what the requesting user could already see. A departed expert's captured judgment doesn't leak the files they once could open.
Human review remains visible
High-risk actions need a licensed human sign-off, recorded in the audit trail. The brain drafts; it doesn't replace the license (19 CFR 111). Approvals are reconstructable end-to-end.
Compliance posture
What's real, targeted, and by design.
EU AI Act
Governance-by-design positions us ahead of full applicability.
GDPR
Tenant isolation, no-egress, customer-held keys, open-format export.
SOC 2 Type II
Roadmap; not yet attained. We won't claim it until the report is issued.
Sector attestations
Pursued per vertical (e.g. HIPAA for insurance). Never blanket-claimed.
Data sovereignty
In-country inference option for restricted entries.
Status as of 2026-06-24. Target / Roadmap = commitments, not claims.
Methodology
How we avoid the overclaim that kills 95% of pilots.
Pre-revenue by choice.
No customer logos to hide behind. Every trust signal on this page is a benchmark, a statute, an architectural commitment, or a guarantee you can verify — not a borrowed name.
Founders & team
To be completed with real credentials before this page ships externally. We won't invent it.
Priced on the outcome, not the headcount.
Value- and outcome-based — aligned to the liability we remove and the senior time we reclaim, not a per-seat search tax. The fine-tuned small model runs at ~1/100 the cost of a frontier call.
~$25–90k/ firm / year
Directional SME pricing — value-anchored to liability avoided and senior-broker hours reclaimed, with a predictable, capped structure for SME budgets.
- Governed, cited brain on one problem lane
- We do the ingestion — read-only, inside your boundary
- Back-test demo on public data, cited, first
- Cited first draft on every case; senior reviews & signs
- Outcome metric agreed up front (senior time reclaimed / error rate)
- Per-firm brain that compounds on your own data
- No per-seat tax · no 100-seat minimum · no cloud lock-in
Outcome-priced · the equation
The metric is agreed up front — senior time reclaimed / error rate — so price tracks the liability we remove, not seats we bill.
vs per-seat search
Glean ≈ $50–65
/ seat · ~100-seat min · retrieves docs, not a cited judgment.
vs frontier API
~1/100 cost
fine-tuned ~70B · runs where your data lives.
The questions a compliance reviewer asks first.
The buying signals we expect — answered directly, not buried in a footer.
See it before it leaves
See what your company already knows — before it walks out the door.
A private diagnostic on a read-only boundary: we map the one problem lane your firm can’t afford to lose, and show you the brain that’s currently walking off the floor.