AI

AI Capabilities
The technology behind the bench

Built on the same models the frontier labs use. Tuned to your practice.

LexEdge is not a wrapper. We choose the right model for each task, ground every output in verified sources, and put a human-in-the-loop on anything sensitive. Eight technical pillars below, in order.

Built with
PyTorch Triton vLLM Llama 3 Mistral Qwen GPT-4 Claude 3.5 Gemini
Case Management
01 pillar 01 of 08

Agentic AI Orchestration

Fourteen specialists. One case file. Clean hand-offs.

Each agent is a constrained reasoner with a single job and an explicit playbook. They share a case-file protocol, pass work forward, and stop at human-in-the-loop gates. Outputs are typed; failures degrade gracefully; every step is logged.

  • Constrained reasoning per agent · narrow tools, narrow prompts
  • Shared case-file schema · typed inputs, typed outputs
  • Human-in-the-loop gates on critical clauses and sensitive matters
  • Full traces · every decision is replayable and exportable
orchestration audit replayable
Case Management
02 pillar 02 of 08

Language Models (LLMs)

Best-of-class models. Switch per matter. Use yours or ours.

LexEdge is model-agnostic. We route per task to the right model — frontier reasoning for negotiation, smaller faster models for drafts and intake. Use OpenAI, Anthropic, Google or run an open-weights model on your own hardware.

  • OpenAI GPT-4 / GPT-4 Turbo · Anthropic Claude 3.5 · Google Gemini
  • Open-weights · Llama 3, Mistral, Qwen for on-prem
  • BYO API key — your costs, your account
  • Per-matter model routing · cost / quality trade-offs you set
multi-model BYO-key self-host-able
Case Management
03 pillar 03 of 08

Fine-tuning · your playbook becomes the model

The bench learns your house style, not the other way round.

Upload a hundred best precedents. We distil your fall-back positions, your boilerplate, your spelling and your firm-specific clauses into a fine-tuned adapter the bench applies automatically. No prompt-engineering required.

  • LoRA adapters per firm · isolated, never shared
  • Built from precedents and red-lines you already have
  • Re-trained as your playbook evolves
  • Versioned and rollback-able · always reversible
LoRA private versioned
Case Management
04 pillar 04 of 08

Retrieval · grounded in real sources

Answers cite a paragraph, not an idea.

Every AI output is grounded in a verified source — the court registry, your matter file, a statute, your playbook. The bench retrieves first, reasons second, and shows you the source so you can defend the answer.

  • Hybrid retrieval · keyword + semantic vector search
  • Re-ranking with a legal-domain reasoner
  • Source-paragraph quotation in every answer
  • No source → no answer. The bench abstains, never invents.
grounded cites-source abstains
Case Management
05 pillar 05 of 08

Vector embeddings · semantic understanding

Search "termination for convenience" — find every clause that means that.

A legal-domain embedding model encodes every clause, judgment and note your firm produces. Synonyms, paraphrases and structurally similar arguments cluster together. You ask the question; the bench knows what counts.

  • Legal-domain embeddings · tuned on Indian case-law
  • Per-firm vector index · permissions enforced at query time
  • Sub-100ms search across 40 years of files
  • Cross-lingual · search in English, find Hindi or Marathi too
semantic cross-lingual fast
Case Management
06 pillar 06 of 08

Voice models · twelve languages

Dictate matters in your mother tongue.

A multilingual speech-to-text model trained for Indian voices, accents and code-switching. Dictate a matter in Marathi, switch to English for the citations, throw in a Hindi name — the bench understands all of it, in real time.

  • Speech-to-text · 12 Indian languages incl. regional accents
  • Code-switching · mid-sentence language changes handled
  • On-device for short utterances · cloud for long dictation
  • Voice commands · navigate the app hands-free
12-langs code-switch on-device
Case Management
07 pillar 07 of 08

Vision models · handwriting & documents

Read a 1989 court diary. Recognise every stamp.

Vision models trained on Indian legal documents — court diaries, deeds, stamped affidavits, hand-written notes, hearing orders. The bench reads them on-device when possible, in the cloud when needed, and reconstructs structured records.

  • Handwriting recognition · English, Hindi, Marathi, regional scripts
  • Document understanding · headers, signatures, stamps, seals
  • Table extraction · cause lists, fee receipts, schedules
  • On-device first · for confidentiality and speed
OCR on-device multilingual
Case Management
08 pillar 08 of 08

Safety · no hallucinations, human in the loop

Confidence is not enough. We need proof.

Every legal AI output runs through a cross-validation gate. The same answer is asked twice with different prompts; sources are checked; citations are matched to the registry. Conflicts surface as flags, not as answers.

  • Cross-validation · two passes, must agree
  • Citation registry check · every authority must exist
  • Human-in-the-loop gates · partner sign-off on critical clauses
  • Refusal-by-default on sensitive matters · escalate to a human
cross-validated human-gate refusal-by-default
The principles
No hallucinations
Sources are checked. Outputs without provenance are refused, not faked.
Human in the loop
Critical clauses, sensitive matters, novel jurisdictions — always a partner's eye.
Your data is yours
Never used for training. Per-firm isolation. Export with one click.

Run the full stack on your own hardware.

For chambers that can\u2019t send a single byte to the cloud, the entire bench runs on a NVIDIA DGX Spark, a Jetson cluster, or a Mac Mini in your office.

See Enterprise Start cloud trial
FAQ
Frequently Asked Questions

Questions lawyers ask before trying LexEdge.

Ten practical answers for this page, marked up with FAQ schema for search engines.
Can I use my own AI model or API key?

Yes. LexEdge is model-agnostic. Firms can bring their own OpenAI, Anthropic or Google API key and pay their own provider costs, or run open-weights models such as Llama 3, Mistral or Qwen on their own hardware. Models can be selected per matter.

How does LexEdge reduce hallucination risk?

Every AI output is grounded in a verified source and runs through a cross-validation gate: the same question is asked twice with different prompts, sources are checked, and citations are matched against the court registry. Where no grounded source exists, the system abstains rather than answering.

Is LexEdge trained on my documents?

No. Firm content is never used to train any model by default. Optional per-firm LoRA adapters are isolated to that firm, never shared, versioned and fully reversible.

Can AI access be restricted by matter?

Yes. Each firm has its own vector index with permissions enforced at query time, so an AI request can only retrieve from matters the requesting user is entitled to see.