All work
demo case studyAI & Agents

Retrieval-augmented support agent

An LLM agent that drafts answers from the company's own docs and hands off when unsure.

Role
AI systems engineer (demo)
Timeline
6-week prototype (demo)
Stack
TypeScript · Next.js · Postgres + pgvector · Drizzle · Open weights / Ollama · Linux
Retrieval-augmented support agent cover
project coverAI & Agents

Demo case study — figures illustrative; client redacted.

Problem

What was broken

A 40-person logistics team spent hours each day answering the same operational questions across Slack and email. Knowledge lived in scattered PDFs and wikis.

Approach

How it was fixed

Chunked and embedded the documentation into a Postgres + pgvector store, then wired an agent that retrieves relevant passages, drafts a cited answer, and escalates to a human when confidence is low. Added a thin evaluation harness so answers could be graded before shipping changes.

Constraints

  • Keep private operational documents inside an open-source-first deployment path.
  • Show supporting passages and escalate instead of inventing an answer.
  • Measure retrieval changes before exposing them to the support team.
Stack

Tools in the system

  • TypeScript
  • Next.js
  • Postgres + pgvector
  • Drizzle
  • Open weights / Ollama
  • Linux
Outcome

What changed

Cut first-response drafting time materially; gave the team a feedback loop to improve coverage week over week. (Figures illustrative — demo case study.)

≈55%Drafting time reduction
< 2 minDraft ready
100%Answers with citations
Lessons

What carries forward

  1. 01Confidence thresholds matter more than a polished chat surface; uncertain answers must stop and hand off.
  2. 02Retrieval quality improves faster when failed questions become evaluation cases, not anecdotal feedback.
  3. 03Citations are part of the product contract: every drafted claim needs a source the operator can inspect.

Related case studies

All work