Skip to main content
TL

AI-assisted commercial offers

Ongoing

AI supports pre-sales, from analysing a tender to drafting the proposal.

adoption after 4 months
70 %
release to production
1 / mois
response-time cut
-25 %
Past documents and offers converge into an AI engine and come out as a structured proposal.

Context

  • Pre-sales works through emails and Excel files, with no central tool.
  • Everyone writes proposals from their own past offers, with no access to the rest of the department's.
  • The team lacks time to build its response strategy for tenders.
  • The project already exists, but it moves at a crawl and rarely reaches production.

My mission

  • Taking over as tech lead, back in production in the first week, then one release a month.
  • Defining the architecture, on the code side and the infrastructure side.
  • Arbitrating technical decisions and reviewing pull requests.
  • Writing the technical documents and the business impact analysis (BIA).
  • Steering two applications and two developers: a senior on the portal, an apprentice on the AI engine.
  • Interface with the business and a say in the makeup of the team.

Decisions and trade-offs

  1. 01

    Bringing vector search into PostgreSQL rather than a managed service

    Problem

    Search runs on a managed service. That service breaks the rules of the mandated LLM provider and costs a lot even at low usage.

    Solution

    Storage and search move to PostgreSQL and pgvector. An external service still produces the embeddings, through a company interface. That interface keeps the data inside the infrastructure.

    Why this choice

    Data governance comes first. Keeping embeddings and search in the in-house database clears the compliance point, a precondition to launching the AI part. Reusing Postgres also avoids the cost of an underused dedicated service. In exchange, the team operates and tunes vector search itself.

  2. 02

    Indexing by chapter rather than by tokens

    Problem

    The first version splits documents by tokens. That split fragments coherent sections and blurs retrieval.

    Solution

    Splitting by business section, with metadata (client, sector, type). Retrieval filters on those criteria first, then ranks by semantic similarity.

    Why this choice

    For the business, a past offer's relevance rests first on structured criteria, ahead of semantic proximity. The structured filter also makes the result explainable and offsets the lack of a dedicated vector index.

  3. 03

    Handing prompts to the business as versioned data

    Problem

    A use case's prompt belongs to the business, yet every change takes a developer and a redeployment.

    Solution

    Prompts live in a versioned store. A senior bid manager edits them in a dedicated interface, tests them, then promotes them to production. The interface keeps every version and restores it in one click.

    Why this choice

    A business autonomous over its own data iterates faster and takes part in quality. Traceability, instant rollback and user training limit the risk of regression.

What this project taught me

  • Taking over a slowed-down project and getting back to production in a week, then on a monthly cadence.
  • Leading each profile differently: light arbitration with a trusted senior, close support for an apprentice.
  • An architecture decision under a governance constraint, as a precondition to launching the AI part.
  • Designing a hybrid retrieval engine, structured and semantic, in a single database.
  • Giving the business autonomy over technical data, with traceability and rollback.
  • Producing the business impact analysis (BIA) and making FinOps calls on the services chosen.

In hindsight

  • Move the long, occasional AI workflows to serverless (Durable Functions). Against response times in seconds, cold start becomes negligible.
  • Tool the corpus pre-ingestion to speed up how it is filled, too slow today.