DiMo

Guiding teams & ideas with leadership, mentorship & vision. Building products that truly serve people.

[DiMo @DiMoFrance]

Envie de rencontrer de nouvelles personnes à #Paris
Discuter en anglais (et un peu en français) autour d’un café, ça vous dit ?
Partagez pour m’aider à créer de belles rencontres !

[DiMo @DiMoFrance]

AsterixLab is live. https://t.co/S4LOs2IcI1.

One correct answer about an OQTF deadline or a titre de sejour renewal can change the direction of someone's life.

I'm building to make that answer reachable. Any hour. Any language. Free.

[DiMo @DiMoFrance]

AsterixLab interroge quatre corpus juridiques en parallèle : https://t.co/7chLAtpLOo (8 langues), le CESEDA - le code régissant tout étranger en France, Legifrance - 156 434 textes en vigueur, et Service-Public - 34 561 fiches pratiques.

203 427 enregistrements. Chaque réponse cite sa source.

[DiMo @DiMoFrance]

The system was live. Then I noticed: every answer cited the same document. No matter what you asked.

Five bugs. All mine. Same-document monopolization. Model anchoring. Cross-corpus duplication. Broken source links.

One week to find them all.

[DiMo @DiMoFrance]

La chose la plus difficile à admettre : j’avais construit quelque chose de techniquement solide, mais fondamentalement faux.

Pas cassé. Faux. Ce n’est pas pareil.

Cassé, on répare. Faux, il faut avoir l’honnêteté de supprimer.

Des mois de travail. Des index FAISS, des modèles locaux, une couche de confidentialité. Tout cohérent. Rien d’aligné avec la vraie question : comment donner une réponse correcte, ce soir, à quelqu’un qui n’a ni les moyens d’un avocat ni le français pour naviguer seul ?

Ce n’est pas un échec technique. C’est une leçon sur ce que signifie construire pour une mission.

[DiMo @DiMoFrance]

Three more corpora. CESEDA - the code governing every foreigner in France. Legifrance - 156,434 articles in force. Service-Public - 34,561 administrative guides.

203,427 records total. All queried in parallel.

29 tests. 29 passing.

[DiMo @DiMoFrance]

9,962 records from https://t.co/FVdfzOVjb0 live. French, English, Arabic, Farsi, Pashto, Russian, Ukrainian, Tigrinya.

1024-dimension vectors. Real retrieval. Answers with cited sources.

For the first time the system did what the mission actually required.

[DiMo @DiMoFrance]

One afternoon in May. D-PIVOT-01 through D-PIVOT-07.

FAISS deleted. BM25 deleted. Ollama deleted. TranslateGemma deleted. Presidio deleted.

Supabase + Mistral-embed + Cloudflare Workers instead.

Starting over is not failure. Choosing wrong is.

[DiMo @DiMoFrance]

#PSGARS 😘

[DiMo @DiMoFrance]

Viva #PSG 🥳

[DiMo @DiMoFrance]

The hardest thing to admit: I had built something technically coherent that was fundamentally wrong for the mission.

Not broken. Wrong.

That distinction matters. Broken you fix. Wrong you delete.

[DiMo @DiMoFrance]

The belief that you need a technical co-founder to build a product is a ghost from a past era.

The constraint has shifted.
It's no longer about who can write the code. It's about who has the sharpest insight into a problem.

We've democratized the 'how'. The 'what' and the 'why' are the only things that matter now.

#productbuilding #saas #startups

[DiMo @DiMoFrance]

7/7
When an agent system keeps reloading large amounts of stable knowledge on every start, the pattern is simple: measure the actual first-turn consumption, build a maintained high-signal distillation artifact, enforce minimal startup plus explicit first-turn discipline, and update the termination process so the artifact stays current.

If you run heavy agent workflows or build your own tooling around these models, what hidden costs have you actually quantified?

[DiMo @DiMoFrance]

1/7
I measured what a fresh agent session actually costs before I ask for anything useful.

New context plus one orientation question consumed roughly 6% of the window. I had not requested real work yet.

[DiMo @DiMoFrance]

6/7
No capability was lost. For any actual creation work, standards changes, or genuinely ambiguous situations, the rules still force full document reads. The model retains complete awareness. The only thing removed was the automatic tax on unchanged context.

[DiMo @DiMoFrance]

First version: local FAISS indices, BM25, Ollama, TranslateGemma, a privacy middleware.

Three experiments. Months of work. It retrieved things. It ran locally.

It could not scale. It could not handle eight languages. I was solving the wrong problem.

[DiMo @DiMoFrance]

Un LLM généraliste va halluciner sur le droit administratif français. Un avocat en France coûte 200 euros l'heure minimum. Un demandeur d'asile avec une échéance et sans français n'a ni l'un ni l'autre.

Ce vide n'est pas un vide de marché. C'est un vide de justice.

C'est pour ça que je construis.