Reconstructed Slides: The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked

Source Video

The Small Model Infrastructure Nobody Built (So We Did) — Filip Makraduli, Superlinked

Method

This deck is reconstructed from the existing video frame captures by detecting likely slide regions with OpenCV, cropping/upscaling those regions, deduplicating similar crops, and OCRing the cropped slide images locally. It is a cleaner companion to the full-stage frame deck.

Reconstructed Slides

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.95
Text source: advanced OCR rapidocr-live/left-72/contrast.
OCR decision: ready — Small embedded article screenshot and composite slide elements are better suited for OCR than direct transcription.

Slide text:

AIE AWRTERSDIAAYONA What Actually Makes Embedding

★ ★ ★ ★ ★ Model InferenceFast?

From Fiash Artention to Quintzatoo,whero is the intergnce borteneck? Is it the architecturo, the mathis,or will wniting ovorything in Rust sohve olmy probiems？hdint:

R's Not Rust)

FIUPVAKRADUUI

N22,2V26

Google Dee

AlEngineer

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.97
Text source: agent_vision.

Slide text:

I realised ... I had to learn

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: agent_vision.

Slide text:

LEARNING FROM FIRST PRINCIPLES

I knew models well

Training & fine-tuning

Evaluation & benchmarking

Applied AI research

vLLM, Flash Attention

I had overlooked infrastructure

How models run in production

GPU utilization & scheduling

Routing & autoscaling

Deployment & operations

Best way to learn?

Join the team building it.

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.96
Text source: advanced OCR rapidocr-live/border-trim/contrast.
OCR decision: ready — Contains a product screenshot with small text plus multiple embedded labels that are better handled by OCR.

Slide text:

SoIjoined theworld class infrastructure team at:

Superlinked github.com/superlinked/sie

AIE VC - backed Al infrastructure company building the model stack for Al search and document processing

Workswithyourfavorite tools

P：Lanceoe OOC 电adamt WWeovoN

erginoering simcle.SIEadcs nstructon-folowng rerankors souopepna dueuogeupue modalraning dacasets and wSEyoucan so-host nfeence for M me reqsred deta transformucior' 'Lenceoe ceotralizes muit- scorng feorrg and ranking host bhem al in one chste comoose the best ndeing modes Weh SE you cen ser- 'Mooern serch sytems uniccis nacural language pre-procnss your outny ag soarchandwnSIEyoucan "Weviee's Query Agene

011022 ceng See AnoeTanycei lco Van Lo C40400-5

AI Engineer

AlEngineer EUROPE

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/bright-screen/opencv-adaptive.
OCR decision: ready — Dense chart and multi-column card slide with small text

Slide text:

Performance dogrades as context flls up. More tokens,worse answers. Contextrofisrea! Manoges what goes INTO context. Context management

★ ★? AIE ★ Tool calling Emboddings; reranking, extroction

CWhy not Just Opus 4.6?n

Works untif contoxt rot kils guolity

Sourco:Chromo Rosqorch

AlEngineer

AlEngineor EUROPE

hcn

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/border-trim/contrast.
OCR decision: ready — Diagram, small labels, and multiple text panels

Slide text:

WHYITMATTERSFORYOURAGENTS

Taxonomy classification 10K categories Shopify product catalogue text + images A production story for tool calling Manages what goes INTO context, Context management

AIE title+image Product sio.extract() sie.encode() NU dosicoton emnbedrot- WSOL Electronics Computers,Laptops Tool calling Embeddings:reranking,extraction

sie.score() roronkcordoates

Switching model type = changing one parameter, not rebuilding infra "Why not just Opus 4.6?" Works until context rot kills quality

Engineering the future of Al

Al Enginecr

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: advanced OCR rapidocr-live/center-82/opencv-adaptive.
OCR decision: ready — Chart with small axes and labels

Slide text:

YIN - MODEL SUPPORTIS KEY, OPEN SOURCEIS NOT STOPPINg

Modcl Pcrformancc vs sizc

AIE ★

★

3:0

Gommo 4 modol porformonco =your infra noods to koepup

AlEngineer

AEnginccy EUROPE

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: none.
OCR decision: ready — Architecture matrix and side notes with small text

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/border-trim/contrast.
OCR decision: ready — Comparison table with many small cells

Slide text:

SAME5STEPS,DIFFERENTIMPLEMENTATION

BERT Qwen2 CoLBERT

1. Norm Post-norm(after) Pre-norm RMSNorm From base arch

2. Q/K/V 3 → we fuse into 1 Separate (GQA) From base arch

AIE

3. Position Absolute lookup RoPE rotation From base arch

4. Attention MHA,allsame GQA (fewer KV) Bidirectioncl

5. Output 1pooledvector 1pooledvector Nvectors(per-token)

Different fromBERT-differentforward pass Differentoutput type entirely

All on packed sequences with flash_attn_varlen = zero padding waste

Engineering the future of Al

AlEnginecr

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: advanced OCR rapidocr-live/center-82/contrast.
OCR decision: ready — Dense promo slide with QR code and small text

Slide text:

ILEARNEDMYLESSON ON INFERENCE

We open sourced both the yin (model inference) and the yang (clust

Superlinked

AIE SlE - Superlinked Inference Engine

Dn-scurce dprioxctiencusterererbedsings,rerankirg.ngertrkscn

B.modelyFrreglrctiors.Frem iactcploDibernetes.AlAeache2.o

github.com/superlinked/sie

Google DeepMind

AlEnginecr

Hidden Non-Slide Evidence

`slide-001.jpg` — sponsor_logo confidence 0.99; sponsor logo wall
`slide-002.jpg` — speaker_stage confidence 0.98; camera shot of speaker on stage with projected slide, not a clean presentation slide

Classification audit: raw/sources/slide-ai-classification/reconstructed/qdh_x-uRs9g/audit.json