Hi, I'm Filip. ML engineer and DevRel, Imperial grad, based in London. This is a short version of what's here. I spent the last few years doing applied engineering, developer relations, and go-to-market growth for inference, retrieval, and agentic execution, and here's some of that work in public.
A lot of it, over the past two years, was for a brand that isn't famous yet. I had to be resourceful from day one, talking my way into new events, organizing, and building. I grew Superlinked's public repo, doubling its stars in a few months, and earned three AI Engineer talks without any sponsorship. That taught me how to make something from nothing, and it does mean all of this was done on hard mode. Put the right backing behind the same energy and it scales and compounds.
I write deep-technical pieces for senior AI engineers on my Substack, the same register your GPU glossary lives in. Start with "What actually makes an embedding model", which is also where my AI Engineer talk started. A companion piece, comparing embedding inference solutions head to head (TEI, Infinity, Fastembed and friends), is the competitive analysis behind the engine choices in section 5.
I'm a co-author on FlashNorm (OpenMachine's transformer-tricks, a NeurIPS/ICLR submission in review), a technique that folds RMSNorm's weights into the next linear layer so normalization overlaps with the matrix multiply. My favourite thing about it is that it's a whole pipeline, not just a paper: I profiled where the time goes, wrote the kernels to prove the win (a Triton kernel and a CUTLASS 3.x SM90 kernel validated on H100, the same Hopper kernel family your Flash-Attention post dissects), measured a +12.77% end-to-end speedup in HuggingFace Transformers, and published the converted checkpoints so anyone could run it.
Then the part I'm proudest of, and the reason it belongs in front of a DevRel team: I turned a research trick into ecosystem adoption. A technique only matters if the engines support it, so the lead author and I ran a campaign to get "weightless RMSNorm" adopted everywhere. I landed the reference PRs in Transformers and vLLM, and my vLLM issue diagnosing why the fused kernel couldn't realize the speedup was precise enough that a community contributor wrote the CUDA kernels to fix it. From there it spread on its own: FlashInfer, Tenstorrent's hardware stack, and others now ship support, and there's an open proposal to bring it into DeepMind's Gemma. That is developer relations aimed at the hardest audience there is, kernel maintainers who only move for measured numbers and runnable code. Evidence first, adoption follows. Same motion your content runs on.
The main one, and the talk I'm happiest about. I did it end to end from scratch: the writing, the directing, reading the audience, the stage presence, all of it tuned to earn views, traction, and brand. Landing the likes, comments, and views it did made me genuinely proud. Watch it here. (Two more AI Engineer talks, from San Francisco, are pending release.)
An honest one: I didn't get to pick the topic or the framing, and had to explain specific integration code. Not how I'd structure a stage talk given the choice, but a solid dive into real code for a technical, retrieval-focused crowd, and salvaging the positives from a constrained brief is its own skill. From 3:28.
A talk on model routing for ODSC West, developed from a hackathon entry (darwin-router, a training-free self-improving router that went 75%→85% accuracy in a day while halving cost per request). It's really the "which model does this agent call?" question, the connective tissue between sandboxes and endpoints, solved the way your platform would: no training runs, adapt at the speed of an index write.
A note on how I speak: I adapt to the event and the crowd, live. I'll say "that makes sense" or "that doesn't" on stage and adjust. The talk serves the room, not the slides.
I built and wrote up the SIE MCP edge: an MCP server that gives any agent (Claude Code, Cursor, claude.ai) document tools (parse, redact, extract, summarize, OCR) where the heavy work runs on open models in your own cloud, so the raw bytes and PII never leave and the model only gets back a small clean artifact (~85% fewer tokens than re-ingesting the document each turn). The writeup is honest about the one hard part, how the bytes have to reach the edge out of band for the privacy claim to hold, which is the kind of caveat I think earns trust rather than losing it.
It's also where the whole thesis you and I care about shows up in miniature: adapt the model when the domain needs it. One experiment fine-tuned a $0.80 LoRA in 41 minutes on a single L4 and lifted in-domain legal retrieval ~18% while holding general performance. Agents need sandboxes to act, endpoints to think, and adapted models to get better, on one platform. Open source, part of superlinked/sie.
All public in superlinked/sie (~3.2k stars). I used your platform as our qualification and benchmarking rig, with your public LLM-engine benchmark as the reference for what normal looks like. Three things we tested:
This testing is also where my Europe feedback (§6) comes from: you feel the US-east control-plane latency from London directly, and you hit the region surcharges and large-payload storage caveat only by setting EU pinning up yourself. Alongside the tests: a rigorous retrieval-ablation study (six pipelines, 2,942 pages, 8,766 graded judgments), production evidence generated on your GPUs (the retail-shelf-audit case), and a partnership track record I'd bring straight over: retrievers merged into LangChain and LlamaIndex plus integration pages across Chroma, Qdrant, Weaviate, LanceDB, and Haystack (explainer video), and co-marketing on partners' own blogs like Qdrant. I also run content as a pipeline, not a heroic effort: repo automations and reusable AI skills that mine our codebase for content, draft from it, and edit their own instructions as we learn what works.
From selling inference into European teams:
I wrote a full EMEA strategy for you: the "trinity" thesis (sandboxes + endpoints + RL/fine-tuning as one self-upselling loop, with the sandbox as the verification layer that keeps agent improvement honest), a London/EMEA community and university ground game, forward-deployed demo blueprints (an agent that learns a customer's repo via sandbox-forked rollouts, with eval-gated LoRA promotion), and a 90-day plan. Happy to walk through it; it's basically the job talk.
Communities: I've co-organized hackathons with AI Tinkerers, {Tech: Europe}, DeepMind, and Alibaba/Qwen (the router came out of the Qwen one), and I'm active across EuroPython, PyData London, ODSC, Haystack, and Berlin Buzzwords. The venues a London office will want, and I arrive with organizer relationships rather than cold emails.
Universities: I've spoken at Imperial (my alma mater) and can open a lecture-plus-workshop circuit at Edinburgh, Oxford, Cambridge, and UCL's DARK lab, seeding the RL groups with credits and starter repos and tying them to ICLR/NeurIPS/EurIPS deadlines so your compute shows up in acknowledgments.
Quant finance: your press already lists unnamed hedge funds as customers, and London is the densest quant hub in Europe (Squarepoint, XTX, G-Research, Man Group, Qube). Bursty research compute, strict isolation, research agents needing sandboxes; they hire from the labs above and buy on in-person credibility, which is the demo-led motion I've closed European clients with.
So, the close: I've built the self-hosted version of what you sell, used your platform in production while doing it, published research and kernels that other people's engines now ship, and talked about this stack at ten-plus conferences across Europe, often at events I helped run. Now I want to do the same thing with a platform that matches the ambition. It's all public, so test any of it.