AI & Guardrails

Where the LLM lives, what it sees, and what it never sees. Pick local, our private API, or bring your own keys.

Where the LLM runs · 3 options

Same product surface, three deployment shapes. You can mix per task — high-sensitivity tabs stay local, lower-stakes drafting can use a stronger frontier model if you opt in.

Local / on-prem
Runs on your hardware. Data never leaves the perimeter.
Model
Ollama · DeepSeek-R1 7B / Llama 3.1 / Mistral
Harvesting
No harvesting
Internet egress
None
Audit
Local, customer-controlled
Best for
Deals under NDA, sovereign data requirements, SOC 2 / GDPR posture.
Pros
  • Zero egress
  • Customer owns the model + weights
  • No per-token billing
Trade-offs
  • Smaller models — slower complex reasoning
  • You manage GPU box
MadLadsLab API
Our private inference at ollama.madladslab.com. We do not train on your data and do not log prompts.
Model
deepseek-r1:7b + SD 1.5 for visual aide
Harvesting
No harvesting
Internet egress
TLS to single endpoint
Audit
Per-request signed receipts, retained 30 days for your audit only
Best for
Teams that want strong privacy without owning hardware.
Pros
  • No harvesting clause in MSA
  • Flat-rate, predictable cost
  • Same API surface as OpenAI
Trade-offs
  • Single-provider dependency
  • Smaller model than frontier
Bring-your-own keys
Your OpenAI / Anthropic / Azure key. Highest quality, you accept that provider's data policy.
Model
GPT-4 class · Claude Opus · etc.
Harvesting
Per provider TOS
Internet egress
Per provider endpoint
Audit
Per provider — we mirror to your audit log
Best for
Tasks where frontier reasoning materially changes the answer.
Pros
  • Best-in-class quality
  • You control the contract
Trade-offs
  • Data leaves perimeter
  • Per-token cost
  • Vendor data policy applies

Guardrail pipeline · ECharts

Every LLM call flows through scrub → policy → schema → router → inference → validator → provenance → signed audit. Click any node for the responsible scope item.

Where the LLM shows up in this pitch

Each Meridian view above has a recommended deployment + the specific guardrails we'd put on it.

View
LLM task
Default deployment
Guardrails
Quality of Revenue
Suggest revenue normalization adjustments + flag anomalies
MLL API
  • PII scrub on customer names
  • Numbers must cite source cell
  • No outbound web search
NWC
Draft peg memo from prior 12 months
Local
  • Memo template enforced
  • Numbers locked to schema
  • Output diffed against actuals before save
Trend & Forecast
Narrate forecast cone + scenario commentary
MLL API
  • No external data injected
  • Caveat block always appended
  • User can swap to BYO key for higher quality
Data Room
Auto-tag documents into folders, extract Q&A from PDFs
Local
  • File hashes logged
  • Reviewer must accept tag
  • Confidence threshold before auto-action
Team Workflow
Daily standup summary, draft IC memo skeleton
BYO (Anthropic)
  • Customer chooses to enable
  • Customer-specific deny-list
  • Audit log mirrored to client

Add to scope

Check items to build your estimate. Totals update live in the cart.

  • One interface, three back-ends. Per-tenant default + per-task override.

    12 hrs Firm $3,600 MLL $900 save $2,700

  • Regex + classifier pre-prompt. Allow/deny rules per role.

    10 hrs Firm $3,000 MLL $750 save $2,250

  • JSON-schema outputs, no free-form leakage, validator rejects malformed.

    8 hrs Firm $2,400 MLL $600 save $1,800

  • Every numeric claim must reference a source cell or document.

    9 hrs Firm $2,700 MLL $675 save $2,025

  • Prompt + response + model + version, signed and timestamped for diligence.

    7 hrs Firm $2,100 MLL $525 save $1,575

  • On-prem install, model management, GPU-or-CPU fallback, healthchecks.

    10 hrs Firm $3,000 MLL $750 save $2,250

  • Golden-set tests so model swaps don't silently change diligence answers.

    8 hrs Firm $2,400 MLL $600 save $1,800