Filip Makraduli

This is a guide to some of my public work. For the last few years I've been doing applied engineering, developer relations, and go-to-market growth at startups working on model inference, search and retrieval, and agentic execution.

Most of it happened at an early-stage startup on its way into hypergrowth. I grew Superlinked's public repo, roughly doubling its stars in a few months, and got 3 AI Engineer talks with zero sponsorship, mostly by building partnerships and creating high-quality technical content. It taught me how to get momentum going from nothing, and honestly I'm excited to see what that looks like with real backing behind it, at a company like Modal.

1. Conference talks and writing

• AI Engineer Europe 2026: "The Small Model Infrastructure Nobody Built (So We Did)." 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.)
• Haystack EU: a longer lecture in three parts (the math one).
  • Part 1, the math and an analysis of DeepMind's paper. I open with Dante and the analogy of Virgil (2:07), then the deep dive (8:28 to ~23:00): sign-rank theory, a satellite analogy, NP-hard problems, and DeepMind's paper on the theoretical limits of retrieval.
  • Part 2 (24:00–30:00): what a mixture of encoders is.
  • Part 3, beyond text-only benchmarks (30:13): evaluating queries in ways the standard benchmarks miss.
• Inference talks (EuroPython, Berlin, Kraków, and more).
  • Prefill and decode for engineers new to inference: 1:49 to 8:45. Prompt processing is compute-bound, token generation is memory-bandwidth-bound, and that changes which hardware and engine you pick.
  • Superlinked's PID-based batching, shared vs non-shared queues, then a competitor-comparison slide and a showcase of integrations and early partnerships: 18:37.
• A Qdrant + LlamaIndex + LangChain talk (Berlin), under tight constraints. 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.
• Coming next (end of October, San Francisco). A talk at ODSC West on hybrid model routing as a way to cut inference cost, developed from a hackathon entry (darwin-router).

A note on how I speak: I adapt to the room, live, and let the talk serve the crowd rather than the slides.

Between talks I write deep-technical pieces for senior AI engineers on my Substack. A few to start with:

• A deep dive on inference and what actually makes an embedding model (also where my AI Engineer talk started).
• An analysis of Flash Attention and the inference story around it.
• A head-to-head comparison of embedding inference solutions (TEI, Infinity, Fastembed and friends), the competitive analysis behind the engine choices in the platform section below.
• The Superlinked × Qdrant demo I wrote on Qdrant's own blog: multimodal hotel search.

2. Code and applied AI

FlashNorm: applied engineering

I'm a co-author on FlashNorm (OpenMachine's transformer-tricks, a NeurIPS/ICLR submission in review). The trick: fold RMSNorm's weights into the next linear layer so the normalization overlaps with the matrix multiply. I figured out where the time was going, wrote the kernels to prove the speedup was real (a Triton kernel and a CUTLASS 3.x SM90 kernel I validated on an H100), got a +12.77% end-to-end speedup in HuggingFace Transformers, and published the converted checkpoints so people could just use it. The notebooks I wrote each check one claim.

FlashNorm: ecosystem growth

A trick is only useful if the engines actually support it, so we ran a little campaign to get weightless-RMSNorm adopted. I landed the reference PRs in Transformers and vLLM, and my vLLM issue explaining why the fused kernel wasn't getting the speedup was clear enough that someone in the community wrote the CUDA kernels to fix it. From there it kept going on its own: FlashInfer, Tenstorrent's stack, and an open proposal for DeepMind's Gemma.

Custom retrievers

I wrote retrievers for LangChain (a published PyPI package) and LlamaIndex on LlamaHub, and explained how they work in this video.

SIE MCP edge

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 actual work runs on open models in your own cloud. So the raw files and PII never leave, and the model just gets back a small clean result (about 85% fewer tokens than feeding it the whole document every turn). And because the models run on your own cluster, you can fine-tune them for a domain: in one experiment a $0.80 LoRA, trained in 41 minutes on a single L4, bumped in-domain legal retrieval by ~18% without hurting the general case. Open source, part of superlinked/sie.

I've also overseen other examples in the repo, like a contract-review agent and an agent-action monitor, but those three are the ones I'd point you to.

3. How I've used your platform

I've tried a few things on your platform: some of the demo notebooks and examples, but mostly I've used it in internal work to test whether specific models run cleanly, using details from your Almanac as the reference. A couple of examples:

• Does a new engine version run cleanly on a given GPU and model? Testing SGLang 0.5.10 with CUDA 12.9 and Qwen3.6-27B on a rented H100, the server booted and loaded fine but a GPU kernel failed to compile on the first real request, because the newer CUDA compiler is stricter. Cheap to catch on an afternoon of H100 time rather than later.
• Does a new kernel path hold up? Checking the FlashAttention-4 beta builds and runs alongside a pinned stack before relying on it. The kernel your team reverse-engineered, tried from the user's side.
• Which GPU for which model? Quick throughput and latency checks across model classes (detection plus OCR runs happily on a cheap L4).

This testing is also where my product learnings come from: from London you feel the US-East latency directly, and you only run into the region surcharges and the large-payload routing once you actually try to pin things to the EU.

Beyond my own tests, I've overseen all of our examples. And I run content as a pipeline rather than a one-off scramble: repo automations and reusable AI skills that mine our codebase for content ideas, draft from them, and update their own instructions as we learn what works.

4. Product learnings and feedback (for you)

Feeding what I pick up (at events, from customers, from benchmarking) back into the roadmap is already how I work. Same instinct, pointed at Modal, from selling inference into European teams:

• The clearest one is data residency. If you send anything above 2 MB it gets routed through US-East, which is a real issue for privacy-heavy and European teams that need their data to stay in region.
• Latency is similar. From Europe you're paying a transatlantic tax because the control plane sits in US-East (around 150ms p95), so having region-local input planes would matter a lot here.
• A smaller one: right now you can swap in a new model version on Modal, but you have to wire it up yourself, run your tests on the new model or fine-tune, only make it live if it beats the current version, and roll back if it doesn't. Baseten and OpenPipe's Loops turned that into a real feature, and as more teams train per-customer models it would be a nice thing to have out of the box.
• A bigger, further-out one: continual learning as something companies can build on Modal. If a team can keep improving their own model from their own data, all of it staying inside their cloud, that is a strong story, and it is an even stronger one in Europe where data retention and private inference really matter. Modal already has the pieces (private compute, fine-tuning, endpoints); packaging them into a safe, keep-your-data loop could be a real wedge here.

5. Strategy, vision, and partnerships

I think the future here is the trinity: Modal owning sandboxes, endpoints, and fine-tunes as one loop, the place where agents act, think, and get better. The natural next step is continual learning on top of that loop, teams improving their own private models over time without their data ever leaving their cloud, which is exactly what European buyers care about. Own the whole loop and the growth is huge, and I've thought a lot about what that looks like in EMEA.

The ground I'd build on, a lot of it warm already: communities and hackathons I've co-organized (AI Tinkerers, {Tech: Europe}, DeepMind, Alibaba/Qwen, plus EuroPython, PyData London, ODSC, Haystack, Berlin Buzzwords); universities and RL labs (Imperial, Edinburgh, Oxford, Cambridge, UCL's DARK lab) through talks, workshops, and credits tied to ICLR/NeurIPS/EurIPS deadlines; the newer AI labs and startups; and quant finance, where London is the densest hub in Europe (Squarepoint, XTX, G-Research, Man Group, Qube) and the workload fits Modal perfectly.