Selected work

Challenge — A growing team needed AI answers grounded in thousands of internal documents, without the hallucination risk of a plain LLM wrapper.

Approach — Built a retrieval pipeline with structure-aware chunking and hybrid semantic + keyword search, so exact-match terms are never missed by vector similarity alone.

Architecture — Document ingestion → structure-aware chunking → embeddings in a vector database, reranked and merged with keyword results before generation. Every answer carries a citation back to its source chunk.

Challenge — Associates were spending hours manually searching contracts and case files for relevant precedent, with no way to verify an AI summary against the source text.

Approach — A citation-grounded research assistant scoped to a firm's own document set, where semantic search surfaces relevant precedent and every answer cites the originating clause or case.

Architecture — Matter-scoped vector indexes keep each client's documents isolated. Clause-level chunking preserves legal structure so citations point to an exact, verifiable passage, not an approximate page range.

Challenge — Engineering teams re-asked questions in chat that were already answered somewhere in a design doc, runbook, or old thread — searchable in theory, findable in practice never.

Approach — An internal assistant grounded in the team's own design docs, runbooks, and incident postmortems, answering in-context inside the tools engineers already use.

Architecture — Continuous ingestion of docs and postmortems into a permission-aware index, with conversational memory for follow-up questions and a fallback to 'not covered in our docs' rather than a guessed answer.

Challenge — A back-office team was manually re-keying data from scanned forms and PDFs into downstream systems — slow, repetitive, and error-prone at volume.

Approach — A layout-aware parsing pipeline that extracts structured data against a configurable schema per document type, instead of treating every document as a flat block of text.

Architecture — Document layout analysis separates text, tables, and figures before extraction; OCR handles scanned input; a schema-driven extraction layer outputs structured JSON to an API consumed by downstream systems.

Challenge — Design engineers needed fast answers to spec and clearance questions without breaking flow to search PDFs across multiple standards documents.

Approach — An agent with typed tool access to the active drawing and a spec knowledge base — geometry checks and standards lookups run as validated function calls, not free-form text generation.

Architecture — A retrieval-grounded spec index feeds an agent that calls typed tools — geometry validation, clause lookup — instead of generating unstructured text, with every tool call logged for review.

Challenge — A growing support team was manually triaging and tagging every inbound ticket before it could be routed, creating a queue backlog during peak hours.

Approach — An event-driven pipeline classifies, tags, and drafts a first-response suggestion for each ticket the moment it arrives, with a human approving before anything sends.

Architecture — Inbound events trigger a classification step, a drafting step grounded in the product's own documentation, and a human-approval gate before any customer-facing action is taken.

Challenge — A legacy client-rendered app had slow first paint and inconsistent behavior across devices, hurting both conversion and SEO.

Approach — Rebuilt on Next.js with server components for data-heavy routes, a shared design system, and a strict performance budget enforced in CI.

Architecture — Server components handle data-heavy routes with zero client JS by default; a shared component library keeps card, typography, and spacing rules consistent across every page.