DigiLawyer

A dropped connection leaves an ambiguous result: did the request fail, or did the response stop reaching the browser? Retrying without resolving that ambiguity can duplicate work.

My work covers idempotent retries, resumable streams and checkpointed execution. These address separate concerns: accepting a request again safely, recovering delivery, and preserving execution progress.

Recovery adds state and failure paths to maintain. The useful distinction is between reconnecting to output and restarting work. Stream recovery alone does not make every action safe to repeat.

An agent workflow may outlive a browser connection or pause for a person. A pause needs enough execution context to continue without treating the response as a new conversation.

I work on LangGraph agent orchestration with resumable human-in-the-loop execution. The engineering focus is execution state: what has completed, what is waiting for input, and what can continue.

Explicit orchestration makes state transitions easier to inspect, but introduces persistence and recovery decisions. A checkpoint is useful only alongside a clear account of which actions have already happened.

A fluent legal answer can still contain an invalid citation or follow the wrong path. A model judging its own output is not enough evidence of correctness.

My evaluation work includes deterministic checks, fail-closed evaluation and model comparisons. Deterministic validation covers properties that can be checked directly; model-based assessment addresses questions that require judgment. In the evaluation harness, malformed judge responses and transport failures produce an inconclusive result, which does not count as a pass.

A failed or unreadable evaluation must not become a passing result. Quality, cost and the limits of each check belong in model comparisons. No single score establishes that a legal workflow is ready to ship.

Legal research needs both exact references and passages that express a related idea in different words. Either kind of match can be missed when search relies on a single signal.

My retrieval work spans PostgreSQL, pgvector, ParadeDB/BM25, citation boosting and reranking. Exact text, semantic similarity and citation signals offer different ways to identify useful sources.

Combining signals introduces ranking choices and more stages to evaluate. Reranking can reorder retrieved candidates; it cannot recover an authority that never entered the candidate set. Sources must remain inspectable alongside the answer.

Connecting an AI client to legal tools introduces an authorization boundary. A tool being discoverable does not mean a caller should be able to use it or access every resource behind it.

I work on authenticated Model Context Protocol (MCP) integration with OAuth and PKCE. The focus is connecting clients to capabilities with an authenticated user context.

A standard protocol improves interoperability while adding consent and credential lifecycle concerns. Authentication establishes who is calling; resource authorization remains a separate responsibility.

A shared AI product has to preserve access boundaries across users and organizations, including when tools retrieve context for a model.

My work includes multi-tenant architecture and authorization. The design concern is carrying access requirements through the full workflow, rather than relying on a model to decide which data a caller may see.

Access boundaries increase the number of retrieval and execution paths that need review. A useful answer and a permitted answer are separate requirements; both need attention during implementation and evaluation.