Half the role runs the platform. Half of it ships on top.
Open weights, self-hosted where it matters. Evaluated before it ships.
The same six steps every time. Step 01 is the one most teams skip.
What I run at home is the same stack I run at work, minus the scale.
plain-text version (screen readers, narrow screens, raw view)
┌─ FIG_001 · WHOAMI ─────────────────────── shell ─┐
│ $ whoami --verbose │
│ │
│ maxime perrot │
│ applied ai engineer at compagnie │
│ financière tradition, one of the │
│ world's largest interdealer brokers. │
│ phd in ai and knowledge graphs. │
│ │
│ $ cat role.split │
│ │
│ [50%] platform │
│ self-hosted llm inference: │
│ model selection and evaluation, │
│ serving optimisation, capacity │
│ planning, ci/cd and mlops. │
│ │
│ [50%] forward deployed │
│ embed with business teams, scope │
│ the problem, build it, ship it, │
│ support and iterate with users. │
│ │
│ $ uptime │
│ │
│ evenings go to the home ai lab. the │
│ rest goes to running, cycling, │
│ swimming and hiking. │
└──────────────────────────────────────────────────┘
┌─ FIG_002 · HOW I BUILD ────────────────── stack ─┐
│ the shape of every system i ship: │
│ │
│ ┌────────────────────────────────────────┐ │
│ │ applications │ │
│ │ chat · agents · tool calling │ │
│ ├────────────────────────────────────────┤ │
│ │ orchestration │ │
│ │ langchain · langgraph · mcp │ │
│ │ retrieval · embeddings · rag │ │
│ ├────────────────────────────────────────┤ │
│ │ evaluation │ │
│ │ llm-as-a-judge · regression │ │
│ │ langfuse tracing │ │
│ ├────────────────────────────────────────┤ │
│ │ serving │ │
│ │ vllm · quantisation · lora │ │
│ │ open weights │ │
│ ├────────────────────────────────────────┤ │
│ │ infrastructure │ │
│ │ kubernetes · gpu · ci/cd │ │
│ │ on-premise or cloud │ │
│ └────────────────────────────────────────┘ │
└──────────────────────────────────────────────────┘
┌─ FIG_003 · THE LOOP ────────────────── delivery ─┐
│ $ cat lifecycle.md │
│ │
│ scope > proto > build > eval > ship > support │
│ ^ │ │
│ └────────────── iterate ───────────────┘ │
│ │
│ 01 scope sit with the team that │
│ has the problem │
│ 02 proto smallest thing that │
│ proves it works │
│ 03 build pipelines, services, │
│ retrieval │
│ 04 eval judged and regression │
│ tested pre-release │
│ 05 ship onto the platform, │
│ behind ci/cd │
│ 06 support monitor, measure, │
│ iterate with users │
│ │
│ domains: front office · risk · ops │
│ regions: americas · emea · mid east │
└──────────────────────────────────────────────────┘
┌─ FIG_004 · HARDWARE ──────────────── nvidia-smi ─┐
│ $ nvidia-smi │
│ │
│ +----------------------------------------+ │
│ | NVIDIA-SMI driver: caffeine | │
│ |----------------------------------------| │
│ | Fleet Memory Where | │
│ |========================================| │
│ | 10+ H200 [##########] work | │
│ | 1 DGX Spark [#######---] home | │
│ +----------------------------------------+ │
│ │
│ at work a cluster of 10+ enterprise │
│ gpus, h200 class included. │
│ every day is a bit like │
│ christmas. │
│ │
│ at home the same stack end to end on │
│ my own hardware: serving, │
│ quantisation, agents, │
│ retrieval. private inference. │
└──────────────────────────────────────────────────┘
┌─ FIG_005 · TRAJECTORY ───────────────── git log ─┐
│ $ git log --graph --oneline --all │
│ │
│ trunk = employment │
│ branch = academic, concurrent │
│ │
│ * apr 2026 -> now │
│ | applied ai engineer, capital markets │
│ | compagnie financière tradition │
│ | · self-hosted llm inference platform │
│ | · forward-deployed delivery, 3 regions │
│ | │
│ * may 2025 -> apr 2026 │
│ |\ founding ai & software engineer, │
│ | | then technical lead │
│ | | scholé ai (epfl spin-off) │
│ | | · prototype to production on azure │
│ | | · hybrid retrieval, agents, vllm/k8s │
│ | | · led 6 engineers; behind $3m raise │
│ | | │
│ | * 2025 -> 2026 │
│ |/ postdoctoral research software eng. │
│ | epfl · llm research-to-production │
│ | · concurrent with scholé ai │
│ | │
│ * oct 2021 -> jan 2025 │
│ |\ phd researcher, ai & data science │
│ | | orisha group, paris │
│ | | · industrial rag over a large │
│ | | knowledge graph; etl at scale; │
│ | | gpu inference on kubernetes │
│ | | │
│ | * 2021 -> 2025 │
│ |/ phd, ai & knowledge graphs │
│ | isae-ensma, poitiers │
│ | · industry-funded, concurrent │
│ | · 2 publications │
│ | │
│ * 2019 -> oct 2021 │
│ |\ product owner & developer │
│ | | ennov, paris │
│ | | · led 5 devs, agile scrum, etmf │
│ | | │
│ | * 2018 -> 2021 │
│ msc computer science, computer │
│ vision · université de poitiers │
│ · honours, merit scholarship │
│ · concurrent with the apprenticeship │
└──────────────────────────────────────────────────┘
┌─ FIG_006 · ELSEWHERE ────────────────── grep -r ─┐
│ languages python · typescript · java │
│ c++ · sql · sparql │
│ services fastapi · django · celery │
│ redis · asyncio │
│ data spark · airflow · kafka │
│ neo4j · postgres/pgvector │
│ ml pytorch · lora · vision │
│ cloud azure (ai foundry, aks) │
│ databricks · aws · gcp │
│ ways agile · hiring & interviews │
└──────────────────────────────────────────────────┘
┌─ FIG_007 · CONTACT ──────────────────────── eof ─┐
│ $ cat contact.txt │
│ │
│ github github.com/maximep2 │
│ linkedin /in/maximeperrot │
│ location lausanne, switzerland │
│ languages français · english │
│ │
│ $ cat DISCLAIMER │
│ │
│ personal account. written in a │
│ personal capacity, my own views, not │
│ my employer's. no proprietary, client │
│ or confidential material appears in │
│ any repository on this profile. │
│ │
│ $ exit │
└──────────────────────────────────────────────────┘