Work / 01 · agentic-tender-assistant
01/ AI agents

Agentic assistant for tender evaluation and drafting

AI pipeline that automated two ends of a heavy documentation cycle: evaluating incoming proposals and assisting the drafting of outgoing specifications.

LangGraphOllama / vLLMEmbedding + ChromaDBRAGMCPPython
01/ Overview

An assistant built on top of an agent graph, operating over a dense and highly regulated body of documents. Manual work by technical analysts on long legal documents went from days to hours, with results that are traceable rather than opaque.

The design is deliberately hybrid: the language model only steps in where it actually adds value: reading ambiguous passages, rewriting free-form text. The heavy lifting (document retrieval, templates, tables, validations) stays in deterministic code, which is auditable, predictable and cheap to maintain.

The whole system runs on-premise. No data leaves the organization, and there is no dependency on external LLM APIs.

02/ Design decisions worth highlighting
  • Predictability over flexibility: the LLM is used in bounded sub-steps, not as a black box that decides what to do.
  • Natural language where it earns its place; classic code where it does not. Fewer surprises, controlled cost.
  • Fully on-premise deployment from day one.