Kidus's Chronicles

Chapter 3 of RAGged But Right is live. This one's about what happens after you deploy, the part most teams don't talk about. Latency that compounds. Knowledge bases that drift silently. Retrieval that looks fine but isn't. And the observability gaps that let it all slide unnoticed. If your RAG system works great in the demo but wobbles in production, this chapter is for you. Link

Dropped my new project: Spokenly. Solid AI app for recording voice, transcribing audio/video, and generating speech from text – everything in one spot. Records clean audio straight in the app, transcribes with timestamps and solid accuracy and does natural TTS with voice options. Handles multiple languages: Amharic, English, Spanish, French, German, Japanese, Chinese, Hindi Built clean with React/Tailwind, and keeps your data private. Good for podcasts, notes, interviews, lectures, whatever. Check it: https://spokenly.kidus-yohannes.engineer Thoughts? Been using any voice AI stuff? What feature would you hit most? Let me know.

I was reading about Graphlit today and honestly… this thing is wild. It’s basically an entire AI content engine in one platform , ingesting videos, PDFs, websites, podcasts, anything , then turning all of it into structured, searchable knowledge. The crazy part? It handles transcription, embeddings, RAG pipelines, content generation, summaries, analytics… all in one place. It’s like someone bundled every tool we usually stitch together and just said “here, use this.” Made me think a lot about how the future of AI won’t be about building everything from scratch, but about using smarter infrastructure that lets you focus on logic, design, and user value instead of plumbing. Anyway, just wanted to share, it’s one of the most interesting platforms I came across today. If you’re into AI, LLMs, or knowledge systems, check it out. https://www.graphlit.com/

Starting the day.

Been Building Agents for a While Now… And man One thing that keeps hurtin, evaluation. It’s not the prompt, not the model it’s measuring if the agent is actually good. and each agent needs a separate metrics, Retrieval Agents → Recall, Depth, Bounce Rate, Context Drift Reasoning Agents → Coherence, Factual Alignment, Self-Revision Planning Agents → Goal Completion, Adaptation Time, Recovery Rate Conversational Agents → Engagement, Tone Adaptation, Helpfulness Memory Agents → Recall Precision, Context Fit, Forgetting Curve Creative Agents → Originality, Constraint Obedience, Iteration Gain That’s when I decided, evaluation isn’t after building. It’s part of building. Without it, agents don’t learn, they just hallucinate confidently. So I’m building something to make evaluation modular, agent-specific, and alive. Because real intelligence isn’t how you generate, it’s how you measure. In the next days i will be sharing everything, Building in public, and solve one of the most vital aspects of AI Engineering

Everyone’s building AI apps with GPT-4… but most still feel dumb. Why? Because the model isn’t the problem, your context pipeline is. 90% of devs just send the user’s message straight to the LLM. But the best AI apps add 3 layers before the prompt ever reaches GPT: 1️⃣ Retrieval: find what’s relevant. 2️⃣ Reasoning: structure that info logically. 3️⃣ Reflection: check if the answer makes sense. That’s how you go from chatbot → co-pilot. Don’t just prompt GPT. Architect it.

Interesting video i came across https://youtu.be/giT0ytynSqg?si=H5qVVJ9QN-bA6DI1

Soooo I’ve got access to Perplexity Enterprise for a while now, which also lets me invite others to try Comet. And after accepting you can get 5 dollars for every person you invite. Although after installing comet and logging in a new account send a one chat request and then paste the link i put again to get perplexity pro 🔗 https://pplx.ai/kidus_yohannes

Hey guys, as you know i made an Ethiopian Earthquake monitor and There are multiple earthquakes happening and you can check using the site i built and there is Sesmic Activity prediction as well as Earthquake risk prediction using Shokribadi Ground motion algorithm and DivS3 optimizations on the site. Check it out and let me know https://earthquake.kidus.engineer/

I have started a linkedin newsletter about AI, ML, ARAG and anything in between. Check out the first chapter https://www.linkedin.com/pulse/ragged-right-chapter-1-foundational-principles-kidus-yohannes-shrvf Edit: I actually posted it as an article not inside a newsletter

My boss sent me this https://x.com/prathkum/status/1971886965818896563?s=46

I was leading the development of one of the projects we will be releasing here. Join the waitlist form. It's going to be exciting.

The interest we have seen so far confirms what we suspected: the challenge of fragmented information is widely felt. Over 200 people across sectors have already joined the waitlist and the list continues to grow each week. This early momentum reinforces a simple truth. The need for reliable, centralized intelligence is urgent, and the people driving Ethiopia’s innovation economy are ready for a solution that meets their growing needs. ⏳ Join the Waitlist if you’re curious https://tally.so/r/w2GRGb

Currently one of the projects i am working on is a scientific platform designed for selecting and scaling ground motion records, which is actually crucial for earthquake engineering and seismic hazard analysis, which is a continuation of the project i worked on before which i have attached above. It is an ML Model which analyzes and processes large seismological datasets, to make reports so that engineers and researchers can match ground motion records to multiple target spectra.

I am migrating from web and app development to ML and AI Engineering, I have been learning for the past few months and doing multiple projects. I even landed my first two upwork clients using that, I will share the details of the resources i used, what I learned and what the next steps are. I know i haven't been posting much, life have gotten so busy, but I will try to be consistent from now on.