L3:ai-solutions-assistant · AI Wing · Featured
Bilingual RAG assistant with citations
A multi-agent assistant for a company AI solutions site, answering only from curated content with citations.
- Role
- Owner and technical lead
- Period
- Jul 2026 to Sep 2026
Results
-23%
Reduction in prompt tokens; model calls per turn fell from 3 to 2
Measured in pull-request benchmarks on a low-traffic site, about 33 turns per day.
Verified~4.2x
Token undercount identified and corrected in the daily spend cap
Identified during a production review.
Verified
Problem
The bilingual company site needed cited answers from curated content, prompt-injection defenses and an enforceable daily budget.
What I did
- I built a deterministic router and three YAML agents using Microsoft Agent Framework.
- I combined Cosmos DB vector search and Azure AI Search hybrid retrieval with citation gating, Prompt Shields and sentence-level output checks.
- I fixed production failures that were not being surfaced: broken stream framing, telemetry reporting zero requests and a daily spend cap that undercounted tokens.
Architecture
- ClientNext.js siteStreaming assistant
- ServiceDeterministic router
- AI3 Agent Framework agents
- AIContent Safety guardsPrompt Shields, output guard
- DataCosmos DB vector and AI Search hybrid
Data flow
- Next.js site to Deterministic router
- Deterministic router to Content Safety guards (screen)
- Deterministic router to 3 Agent Framework agents
- 3 Agent Framework agents to Cosmos DB vector and AI Search hybrid (cited passages)
Tech stack
Next.js
Microsoft Agent Framework
Cosmos DB
Azure AI Search
Azure OpenAI
Azure AI Content Safety
App Service
L2 · Related career entryTechnical Consultant, Software & AI · Metrodata (PT Mitra Integrasi Informatika)