AI at Francium

Engineers steer. AI ships.

We build with AI every day, and we build AI into the products we ship for clients. Engineers decide what to build, in what order and with what tradeoffs. The AI does the heavy lifting. Same reviews, same CI, more ambition.

The core principle

Context is
the multiplier.

The quality of what an AI builds rises with the context it is given. A one-line prompt gets a generic answer. A full brief gets work that fits.

  1. Minimal prompt
  2. + Docs
  3. + Codebase
  4. + UI and history
  5. Full context
Output quality rises with each layer of context provided.

1Mtokens in a single conversation

What full context means.

Everything an engineer would want before touching the code, loaded into one conversation the AI holds in working memory.

  • Requirement documents
  • The codebase
  • Screenshots
  • UI mockups
  • Git history
  • Test output
  • Review comments
  • Stakeholder feedback

Same task, two prompts

A one-liner
versus a brief.

A real task: untangling a long-forked open-source dependency. The ask is the same. The context is not.

Shallow

Fix the systemic issue with the forked dependency. Propose a plan.

The AI does not know what the dependency is, why it is a problem or what you prefer. It answers with a generic audit.

Deep

There is a systemic problem in this project: a dependency forked from an open-source project years ago. Forking froze it in time, and a lot of work has been done on top since.

Find which upstream version it was forked from and every change applied on top. Understand what each one does and how it is used today. Be exhaustive.

Then we will weigh options, and you can propose your own. The ones on my mind:

  • Rewrite what it does as a small, independent service.
  • Rewrite it as a fresh project with no open-source base.
  • Move to the maintained community successor and carry our changes as a thin layer on top.

Think like a senior architect. A project like this should never be forked; at most, patch it in place and cover the patch with tests, so upstream can move and the tests still guard our changes.

Tell me if any of this does not make sense.

The AI now knows the problem, the architecture philosophy, three concrete options and the quality bar. One prompt, a thorough analysis and a recommendation.

The feedback loop

Measured in minutes,
not weeks.

In the time a traditional cycle takes to come back with feedback, the AI-assisted loop has gone round again and again. Each turn is small, and each one is seen.

TraditionalDays to weeks

  1. Spec
  2. Ticket
  3. Design
  4. Build
  5. Pull request
  6. Feedback

AI-assistedMinutes

  1. Describe
  2. Build
  3. See
  4. Adjust
  5. Repeat
Traditional: Spec, Ticket, Design, Build, Pull request, Feedback, days to weeks. AI-assisted: Describe, Build, See, Adjust, minutes, many times over in the same span.

700+iterations on the clickable prototype alone

This is not one prompt and one answer. It is a conversation: paragraphs of context, screenshots of what looks wrong, a correction that takes seconds, repeated until it is right.

The approach

Explore, plan,
then build.

Exploration is kept apart from implementation, so feedback arrives before production code is written.

  1. 01

    Explore

    Feedback in minutes

    The AI builds a throwaway, clickable prototype straight from the requirements. We host it and stakeholders click through it, so everyone reacts to the same thing.

  2. 02

    Plan

    Engineer-driven

    Engineers split the work into release phases by technical dependency, so it ships to production in increments, not in one leap.

  3. 03

    Build

    Same PR process

    Now the AI has full context: documents, prototype, feedback, codebase and review history. Production code follows, reviewed like any other.

Code review as context

Seven comments,
one session.

Review comments are context too. On one pull request every comment went to the AI at once, and a single refactor addressed all of them.

  1. 01Leave the existing feature flags untouched.
  2. 02Scope every query to the device.
  3. 03Pick the data source from the customer's plan.
  4. 04Isolate the new feature under its own route.
  5. 05Move hardcoded test addresses into settings.
  6. 06Confirm the licence of the map tiles.
  7. 07Add a per-customer toggle to the admin form.

One refactor session

All seven, addressed together.

Then back through the same review, like any other change.

The foundation

Make the project
AI-ready.

A one-time investment that pays off on every feature after it. If the AI cannot start, run and test the project on its own, the engineer becomes the bottleneck.

your-project/

  • Makefile

    One-command startup

    One command starts the full stack, so the AI can run what it builds.

  • AGENTS.md

    An agent brief

    Architecture, commands and conventions, read at the start of every session.

  • tests/

    Fast tests

    The AI writes, then verifies. Seconds of feedback, not minutes.

  • PROGRESS.md

    Persistent context

    Progress and strategy notes that survive between sessions, so knowledge compounds.

How it played out

An enterprise security platform.

From 40+ requirement documents to production, in four phases, through the same review process as any other feature.

  1. Phase 1

    Clickable prototype

    8 pages from 40+ requirement documents, hosted for stakeholder feedback.

  2. Phase 2

    Core platform

    Dashboard, incidents and assets, with live detection.

  3. Phase 3

    Advanced detection

    More detection engines, baseline profiling, asset detail and the attack storyline.

  4. Phase 4

    Response actions

    Waiting on firewall hardware.

The only blocker was external hardware.

Everything that could be built with code got built. Same pull requests, same reviews, same CI pipeline.

AI in your product

And we build AI
into products.

The same engineers put models into client products. We use the tool the problem needs, measure it against a plain baseline and know the cost of every request.

  1. 01

    Machine learning

    Forecasting, segmentation, scoring and classification on your own data. Often gradient-boosted trees and good features, sometimes deep learning, always measured against a plain baseline.

    Python · scikit-learn · TensorFlow · PyTorch

  2. 02

    Computer vision

    Image classification, detection and inspection, plus document understanding: reading scans, forms and invoices and turning them into structured data a system can trust.

    Detection · OCR · Document extraction

  3. 03

    Natural language processing

    Classification, entity extraction, sentiment and search over text. Small task-specific models where they are cheaper and more predictable than a general one.

    Text classification · NER · Search

  4. 04

    LLM-backed features

    Retrieval over your own documents, summarisation, drafting and assistants inside existing products. Built with evaluation sets, guardrails and a known cost per request.

    RAG · Tool use · Evaluation · Guardrails

  5. 05

    Data pipelines and MLOps

    Ingestion, feature and training pipelines, model deployment and monitoring. The unglamorous plumbing that decides whether a model still works in six months.

    Batch and streaming · Model serving · Monitoring

Use cases

Where it earns its keep.

Problems we take on, with a note where we have already shipped one.

  • Document processing

    Parse, classify and extract from documents at volume, with a human in the loop for low-confidence cases.

    ShippedA document pipeline with an embedded ML step, taken from about 15 minutes per document, one at a time, to under a minute with no limit on parallel jobs.

  • Customer segmentation

    Group customers by behaviour and purchase history so promotions, pricing and outreach fit each segment.

    ShippedRetail segmentation on a year of purchase history with Google Dataflow and TensorFlow.

  • Search over your own knowledge

    Answer questions from manuals, policies, tickets and wikis, with citations back to the source paragraph.

  • Web data at scale

    Collect and normalise public data from many sources, then find the patterns that matter to the business.

    ShippedConcurrent collection across tens of thousands of websites, with monitoring and a clean service layer.

  • Support and operations assistants

    Draft replies, triage tickets and summarise long threads inside the tools your team already uses.

  • Forecasting

    Demand, load and risk forecasts that planners can read, question and override.

Bring us the ambitious version.

Tell us what you would build if the feedback loop took minutes. An engineer will reply.