AI systems, data engineering and software development that hold up in production.

Problem

Repair steps and safety cautions have to come from the manual for the exact vehicle, not from a model's general knowledge. The manuals are long scanned PDFs, and their diagrams are labeled only by figure number.

Approach

A vision model parses every page into procedures, tables, cautions and figures. Passages and cropped figures are embedded into a Postgres vector index, so a plain question can retrieve the right diagram, filtered to the exact vehicle and returned with structured citations.

Result

In production for a library of manuals of up to about a thousand pages each, answering lookups in milliseconds. Questions the manuals don't cover return nothing instead of a guess.

Problem

Support agents needed the right policy, pricing or how-to answer for every question. Plain vector search kept returning answers that sounded similar but were off-topic.

Approach

Documents and questions are embedded separately and stored in Postgres with pgvector. One query blends vector similarity, full-text rank and keyword overlap into a single confidence score, and a corroboration check rejects close matches that share no words with the question.

Result

In production behind the company's chat and email support agents, covering a few hundred answers across about fifteen categories, with a full-text fallback if the embedding service is down.

Problem

The previous chatbot was hard to maintain, could not take actions safely and charged a fee per turn. Most account questions ended in a hand-off that cost customers a day or two.

Approach

A self-hosted agent with specialist sub-agents reads data through parameterized tools and writes only through backend endpoints. Sign-in checks, preview-then-confirm on every change and an output filter for personal data run in code, and every release is scored on replayed conversations.

Result

Live in production, answering in about four seconds per turn at roughly an eighth of the previous platform's cost per turn.

Problem

Clinical event streams arrive late, out of order or duplicated, and analytics teams need data they can use without exposing patients.

Approach

Event-driven pipelines on Kafka and Pub/Sub, processed with Flink SQL and Dataflow into BigQuery. Watermarks handle late events, exactly-once processing removes duplicates, sensitive fields are de-identified at scale, and column-level security controls who sees what.

Result

Tens of millions of records de-identified on their way to analytics, with access controlled down to the column and lineage tracked end to end.

Problem

Projects, time logs, resource plans, margins and staffing lived in separate tools, so managers had no single view of project health.

Approach

A React and TypeScript web app on a serverless Firebase backend, with role-based security covered by hundreds of automated rule tests. Built-in AI returns structured results: a bilingual assistant grounded in each person's own work, plan suggestions, risk views and staffing matches.

Result

In production as the firm's system for projects, profitability and staffing.