Technical reports on the models and hardware we test, and blog posts about the systems we build. https://lokahq.github.io/tech-blog/ 2026-10-05T00:00:00.000Z https://lokahq.github.io/tech-blog/running-large-language-models-fully-offline-on-mobile-with-react-native/2026-10-05T00:00:00.000Z2026-10-05T00:00:00.000Z

Paolo Pecis explores fully offline speech-to-text and LLM summarization on iOS and Android with React Native, whisper.rn, llama.rn, and the SoloAI app.

Paolo Pecishttps://lokahq.github.io/tech-blog/authors/paolo-pecis/ https://lokahq.github.io/tech-blog/agentcore-identity/2026-09-22T00:00:00.000Z2026-09-30T00:00:00.000Z

Authentication and authorization for AI agents on AWS. Who is allowed to call your agent, and how your agent proves itself to everything it calls.

Matheus Diashttps://github.com/maledias https://lokahq.github.io/tech-blog/esmc-trainium2/2026-08-18T00:00:00.000Z2026-09-30T00:00:00.000Z

ESMC-300M scored 490.8 protein variants per second on one trn2.3xlarge at batch size 16, with ROC-AUC 0.8525 and estimated cost of $1.26 per million variants on this fixed-shape D2Deep evaluation.

João Correiahttps://lokahq.github.io/tech-blog/authors/joao-correia/Telmo Felgueirahttps://lokahq.github.io/tech-blog/authors/telmo-felgueira/Tiago Gonçalveshttps://lokahq.github.io/tech-blog/authors/tiago-goncalves/Bojan Jakimovskihttps://github.com/ShekswessJim Burtofthttps://lokahq.github.io/tech-blog/authors/jim-burtoft/Louise Pinghttps://lokahq.github.io/tech-blog/authors/louise-ping/ https://lokahq.github.io/tech-blog/a-prompt-pointed-the-other-way/2026-08-04T00:00:00.000Z2026-09-30T00:00:00.000Z

Everything is buildable now. Every team I meet lives in the same climate: more ideas than quarters to test them in, more tools than problems, roadmaps that read like wish lists because nothing on them is technically…

Ana Markovićhttps://medium.com/@ana.markovic https://lokahq.github.io/tech-blog/benchmarking-grok-4-3/2026-07-10T00:00:00.000Z2026-09-30T00:00:00.000Z

A sequential benchmark of Grok 4.3 on Bedrock Mantle and the xAI API, covering all four reasoning-effort settings on the full GSM8K test split.

Nina Cvetkovskahttps://github.com/NineCvetBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/grok-bedrock-xai/2026-07-10T00:00:00.000Z2026-09-30T00:00:00.000Z

On 1,319 GSM8K test questions, Grok 4.3 reached 96.51% accuracy through Bedrock Mantle and 96.89% through the xAI API at low reasoning effort; provider differences were clearer in tail latency than answer accuracy.

Nina Cvetkovskahttps://github.com/NineCvetBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/contributing-to-open-source-bioinformatics-our-experience-at-the-nf-core-hackathon-from-medellin/2026-07-09T00:00:00.000Z2026-09-30T00:00:00.000Z

If you work anywhere near data-heavy science — genomics, metagenomics, proteomics — you’ve probably hit the same wall: building reliable, reproducible data pipelines is hard. Nextflow is an open-source workflow engine…

Andres Florianhttps://github.com/TheGreatJackDaniel Sabogalhttps://github.com/daasabogalro https://lokahq.github.io/tech-blog/from-proteins-to-pipelines-how-open-source-nextflow-tools-are-accelerating-drug-discovery/2026-07-08T00:00:00.000Z2026-09-30T00:00:00.000Z

Modern computational drug discovery depends on complex workflows that connect AI models, structural biology tools, databases, and quality-control steps into reproducible pipelines. While many organizations have access…

Jelena Pejovichttps://medium.com/@jelena_82604Jorge Moura Sampaiohttps://github.com/JorgeMouraSampaio1 https://lokahq.github.io/tech-blog/computer-vision-at-scale-on-sagemaker-jobs-four-decisions/2026-07-06T00:00:00.000Z2026-09-30T00:00:00.000Z

Training a vision model on millions of images is, more than anything, a data problem wearing an ML costume. The architecture can many times be the least interesting part. Everything around it is usually more fun: how we…

Didierhttps://github.com/PedroDidier https://lokahq.github.io/tech-blog/pushing-open-source-tts-models-to-their-limits-six-paradigms-14-models-one-production-reality/2026-07-02T00:00:00.000Z2026-07-07T00:00:00.000Z

In the digital audio subset of AI, the questions currently on everyone’s mind are

Ervin Shaqirihttps://github.com/ervin-shAlexandre Domingueshttps://github.com/Taxuspt https://lokahq.github.io/tech-blog/aws-bedrock-guardrails-implementing-an-allowlist/2026-07-01T00:00:00.000Z2026-09-30T00:00:00.000Z

When building AI agents with AWS Bedrock, guardrails are your first line of defense for keeping conversations on topic. A common requirement is to restrict an agent to a pre-defined strict allowlist, blocking everything…

Guilherme Ribeirohttps://github.com/ribeiro1505 https://lokahq.github.io/tech-blog/tts-emotion-benchmark/2026-07-01T00:00:00.000Z2026-07-01T00:00:00.000Z

Fourteen open-source text-to-speech models across six emotion-control paradigms, benchmarked against AWS Nova Sonic v2 on naturalness, expressiveness, latency, and licensing.

Ervin Shaqirihttps://github.com/ervin-shAlexandre Domingueshttps://github.com/Taxuspt https://lokahq.github.io/tech-blog/running-six-open-source-cofolding-models-on-aws-lessons-learned-from-a-compute-perspective/2026-06-30T00:00:00.000Z2026-09-30T00:00:00.000Z

A compute benchmark of six open-source cofolding models on AWS, measuring runtime, memory, and estimated cost across 29 PROTAC ternary complexes.

Julián F. Fernándezhttps://github.com/jf-fernandezAndres Florianhttps://github.com/TheGreatJackDaniel Sabogalhttps://github.com/daasabogalro https://lokahq.github.io/tech-blog/watching-a-python-to-rust-rewrite-was-painful-enough-to-build-this/2026-06-26T00:00:00.000Z2026-09-30T00:00:00.000Z

How we built an agent that ports Python ML code to edge hardware for space.

João Afonso Pereirahttps://github.com/joao-afonso-loka https://lokahq.github.io/tech-blog/benchmarking-gpt-5-5-on-amazon-bedrock-vs-the-openai-api-the-engineering-walkthrough/2026-06-15T00:00:00.000Z2026-09-30T00:00:00.000Z

A reproducible, sequential comparison of GPT-5.5 on Bedrock Mantle and the OpenAI API across 1,319 GSM8K questions and three reasoning-effort settings.

Petar Kalinovskihttps://github.com/PetarKalinovskiBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/openai-bedrock/2026-06-15T00:00:00.000Z2026-09-30T00:00:00.000Z

On 1,319 GSM8K questions, GPT-5.5 accuracy differed by at most 0.22 percentage points between Bedrock Mantle and the OpenAI API; Bedrock had lower time-to-first-byte and higher measured output throughput in the June 2026 sequential test.

Petar Kalinovskihttps://github.com/petarkalinovskiBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/carbon-trainium2/2026-05-21T00:00:00.000Z2026-09-30T00:00:00.000Z

Carbon-500M, Carbon-3B, and Carbon-8B compiled and ran through NxD Inference on one trn2.3xlarge; the benchmark reports fixed-shape throughput, DNA-suite output, and a separately qualified A100 reference comparison.

Bojan Jakimovskihttps://github.com/ShekswessLoka Applied Researchhttps://www.loka.com/ https://lokahq.github.io/tech-blog/running-hugging-faces-carbon-on-aws-trainium2-with-nxd-inference/2026-05-21T00:00:00.000Z2026-06-14T00:00:00.000Z

Hugging Face Bio’s Carbon release is interesting for two reasons at once. First, it is a biology model. Second, it is not an infrastructure outlier. Many genomic models come with custom architectures, specialized…

Bojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/data-governance-in-practice/2026-05-04T00:00:00.000Z2026-09-30T00:00:00.000Z

Lineage for a CDC Pipeline on AWS

Cecilia Brusquettihttps://github.com/cecibrus https://lokahq.github.io/tech-blog/beyond-static-rules-building-agentic-content-evaluation-systems-on-amazon-bedrock/2026-04-22T00:00:00.000Z2026-09-30T00:00:00.000Z

How to turn scattered guidelines into adaptive workflows that score, explain, and improve content at scale

Crhistian Cardonahttps://medium.com/@crhisto https://lokahq.github.io/tech-blog/ia-comedia-a-loka-made-lab-for-theatre-with-live-ai-actors/2026-03-25T00:00:00.000Z2026-09-30T00:00:00.000Z

A Portuguese playwright wanted to push AI beyond the assistant role and onto the theatre stage. Here’s how LOKA built a live system that brings a cast of AI actors to perform alongside humans.

Andreia Pereirahttps://github.com/decasppereira https://lokahq.github.io/tech-blog/the-agent-assembly-line/2026-03-02T00:00:00.000Z2026-09-30T00:00:00.000Z

Productionizing Agentic Use Cases in Weeks, Not Months

Mario Petkoskihttps://github.com/mariopetkoskiBojan Jakimovskihttps://github.com/ShekswessZafir Stojanovskihttps://github.com/zafstojano https://lokahq.github.io/tech-blog/deploying-trinity-mini-drugprot-think-on-amazon-sagemaker-ai/2026-02-23T00:00:00.000Z2026-09-30T00:00:00.000Z

If you work in regulated domains (Healthcare, Life Sciences, Finance) you routinely hit constraints that break the default “just call a hosted API” approach:

Bojan Jakimovskihttps://github.com/ShekswessPetar Kalinovskihttps://github.com/PetarKalinovski https://lokahq.github.io/tech-blog/trinity-drugprot/2026-02-23T00:00:00.000Z2026-09-30T00:00:00.000Z

Twelve sequential ablations on Trinity Mini show that LoRA alpha 64, learning rate 3e-6, and a 2,048-token generation budget formed the most stable effective recipe in this DrugProt relation-extraction study.

Bojan Jakimovskihttps://github.com/ShekswessPetar Kalinovskihttps://github.com/petarkalinovski https://lokahq.github.io/tech-blog/clean-architecture-with-strategy-pattern-in-kotlin/2026-01-27T00:00:00.000Z2026-01-27T00:00:00.000Z

When building stable and maintainable software, one of the biggest challenges developers face is that modern applications often need to support multiple versions of a feature—different algorithms or formats that can…

Stefanija Zdraveskahttps://github.com/StefanijaZ https://lokahq.github.io/tech-blog/building-rag-systems-on-aws-lessons-from-serverless-and-ec2-benchmarks/2026-01-20T00:00:00.000Z2026-09-30T00:00:00.000Z

A practical comparison of RAG ingestion and retrieval architectures using in-memory vector stores across AWS Lambda, EC2, and local environments.

Romulo Pagnozzihttps://lokahq.github.io/tech-blog/authors/romulo-pagnozzi/Crhistian Cardonahttps://medium.com/@crhisto https://lokahq.github.io/tech-blog/rethinking-bioinformatics/2026-01-05T00:00:00.000Z2026-09-30T00:00:00.000Z

Solving the Challenges of Traditional Pipelines

Sara Oquendohttps://medium.com/@sara.oquendo https://lokahq.github.io/tech-blog/part-2-agentic-patterns-101-with-loka-strands-agents/2025-09-02T00:00:00.000Z2026-09-30T00:00:00.000Z

After exploring the foundations of Agentic Patterns in Part 1, where we looked at more structured workflows like Sequential and Parallel workflows, LLM Routing, and Reflection, this time we are going to push further…

Nina Cvetkovskahttps://github.com/NineCvetPetar Kalinovskihttps://github.com/PetarKalinovskiBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/part-1-agentic-patterns-101-with-loka-strands-agents/2025-08-27T00:00:00.000Z2026-09-30T00:00:00.000Z

These days, Agents are everywhere in daily conversation and it seems everyone wants one. We read or hear about AI Agents, Agentic AI, Agentic Architectures, and Agentic Patterns almost every day. But what do these terms…

Nina Cvetkovskahttps://github.com/NineCvetPetar Kalinovskihttps://github.com/PetarKalinovskiBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/flattening-the-curve-with-nextflow-building-a-scalable-and-reproducible-bioinformatic-workflow-for/2025-06-28T00:00:00.000Z2026-09-30T00:00:00.000Z

Loka’s participation in the recent nf-core hackathon in Medellín, Colombia, culminated in the successful development of MITNANEX, a pipeline designed to extract and assemble mitochondrial genomes, identify and annotate…

Jelena Pejovichttps://medium.com/@jelena_82604José Ribónhttps://github.com/joseribonAndrés Sacrehttps://lokahq.github.io/tech-blog/authors/andres-sacre/Federico Ruedahttps://github.com/federuedaNicolás Francohttps://github.com/ndf1511Javier Dominguezhttps://github.com/javi-domiJuliana Silvahttps://lokahq.github.io/tech-blog/authors/juliana-silva/ https://lokahq.github.io/tech-blog/how-to-transition-from-xml-android-view-to-jetpack-compose-view/2025-06-23T00:00:00.000Z2026-09-30T00:00:00.000Z

Moving from XML to Jetpack Compose helps developers write cleaner, faster and more modern UI code — essential for staying up to date with Android’s best practices.

Gabriel Menezes da Silvahttps://github.com/ggabriel-loka https://lokahq.github.io/tech-blog/aws-s3-browser-will-save-you-time/2025-06-17T00:00:00.000Z2026-09-30T00:00:00.000Z

Simple interface for data stored in S3

Dragan Savevskihttps://github.com/dragansavevski https://lokahq.github.io/tech-blog/part-4-benchmarking-deepseek-cost-performance-and-business-use-cases-on-aws/2025-04-10T00:00:00.000Z2026-09-30T00:00:00.000Z

A small GSM8K latency benchmark comparing DeepSeek-R1 Distill Llama 8B deployment paths on Bedrock, SageMaker, and Inferentia2 EC2.

Crhistian Cardonahttps://medium.com/@crhistoBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/part-2-deploying-distiled-deepseek-r1-models-on-amazon-sagemaker-ai/2025-02-20T00:00:00.000Z2026-09-30T00:00:00.000Z

A deployment walkthrough for serving distilled DeepSeek-R1 models with Amazon SageMaker AI, including the AWS setup for open-weight reasoning models.

Crhistian Cardonahttps://medium.com/@crhistoBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/part-3-deploying-deepseek-r1-models-on-aws-designed-silicon-instances/2025-02-20T00:00:00.000Z2026-09-30T00:00:00.000Z

Engineering notes on deploying DeepSeek-R1 models to AWS-designed silicon instances and serving open-weight reasoning models on AWS.

Crhistian Cardonahttps://medium.com/@crhistoBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/harnessing-open-source-ai-on-aws/2025-02-17T00:00:00.000Z2026-09-30T00:00:00.000Z

Deploy distilled DeepSeek-R1 open-weight models through Amazon Bedrock, with a hands-on AWS setup from the Loka engineering team.

Crhistian Cardonahttps://medium.com/@crhistoBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/dive-into-deepseek/2025-02-12T00:00:00.000Z2026-09-30T00:00:00.000Z

How the open-recipe LLM is transforming GenAI

Crhistian Cardonahttps://medium.com/@crhistoBojan Jakimovskihttps://github.com/Shekswess https://lokahq.github.io/tech-blog/how-http-3-0-and-quic-solve-head-of-line-blocking/2025-01-20T00:00:00.000Z2026-09-30T00:00:00.000Z

You probably have heard the history of how HTTP was created and you might even know about the new HTTP/3 version. In this article, I want to focus on some of the lesser-known details of HTTP and why we need a new…

Daniel Pereirahttps://medium.com/@daniel.pereira_7655