Dmytro Palaniichuk 🇺🇦

Head of Engineering @promova_app | Ocean 🌊 lover 🏄‍♂️ , 🐶 🐱 owner

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

At 1 a.m. I asked an Claude to build me a camera with no AI in it.

34 minutes later: an instant camera in the browser. Zero image files, zero audio files. Every click, flash whine and motor whirr is synthesised from code
The print takes 20 seconds to develop. You cannot skip it

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

Before the next 1:1 the previous notes come back as a short brief: what we agreed, what's still open.

The video is code too. Opus wrote a render(t) function, Chrome captured the frames, ffmpeg + a tiny synth did the rest. No editor. #localllm #ollama #anarlog #gemma

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

My 1:1s now turn into .MD notes + my action items without me typing a word.

Recording, transcription and summary all run on my Mac. $0. Nothing leaves the laptop.

Anarlog → Parakeet → Gemma 3:4B (Ollama) → .md

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

Cheap quality check: every note records its % of Cyrillic. Below 70% = something broke, almost always context.

Honest limits: names get mangled, and with 3+ people everyone on the far side collapses into one "them" (mic = me, system audio = them)

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@denicmarko only htmx we trust

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@vercel_dev @oradotai https://t.co/EitLvHir92 scores 89/100 for AI-agent readiness. Good shot @promova_app 🤙

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@bcherny @swyx Yes , would be nice , now we using orca ide wtoh build in cron to clean up old worktree

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

We run both Cowork and Claude Code daily. One frustrating gap: Cowork artifacts can only share org-wide or not at all no picking specific people. CC already does per-person sharing.@Claude, any chance Cowork gets the same control? Small fix, real workflow difference. @noahzweben

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@jerryjliu0 Two years of indexing infrastructure work is exactly the kind of unsexy foundation that determines whether RAG actually ships to production or stays a demo. We learned that lesson the hard way — retrieval quality lives in the pipeline, not the model.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@cognition MapReduce framing for multi-agent security scanning makes sense — decompose by file/module, reduce to deduplicated findings. Curious how it handles cross-module vulnerabilities where the attack surface only appears when two distant components are read together.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@LangChain @pipecat_ai Voice tracing end-to-end is the part that usually breaks first. We built a voice layer on top of an existing agent graph and the hardest debug was always "did the STT mis-transcribe or did the LLM hallucinate?" — without unified traces those look identical in the logs.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@levelsio The upstream rate limit reveal is the real lesson here. OpenRouter abstracts the provider until it doesn't — then you're debugging someone else's quota with no visibility. Direct keys cost more setup but you own the ceiling.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@_philschmid The credential interception pattern is clever but flips the trust boundary — now your Gemini API layer is the thing that must never be compromised, not the sandbox. We found that boundary shift harder to audit than the original problem.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@jerryjliu0 Matches what we found — we spent months chasing novel retrieval architectures, then realized our recall issues traced back to chunking strategy and metadata quality. Boring fixes, real gains.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@jerryjliu0 The filesystem
metaphor is the right abstraction. Moving from "query the index" to "navigate the knowledge base" made agents far more
predictable — they scope reads before writing, like a good engineer.

Partial-update primitives, or is re-index always
full corpus?

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@simonw We felt this acutely during onboarding. New engineers would ship features faster with the agent but couldn't debug production incidents — they'd never built the mental model of what they were deploying. Fast output, brittle understanding.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@Sergiomasellis @BraceSproul It make sense , as soon we’ve faced same issue to sync client/server tech doc for retrieval, still figuring out how to do it better

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@svpino State persistence is where most of our early agent attempts died. Step 1 would finish, get interrupted, and restart from scratch — burning tokens and confusing the user.

What worked: treat every state transition as a write-first operation, not a side effect.

[Dmytro Palaniichuk 🇺🇦 @DPalaniichuk]

@hwchase17 We've run stateful agent evals internally and the hardest part isn't the framework — it's defining what "correct" even means across a multi-step session. Did you land on outcome-only scoring or are you capturing intermediate state fidelity too?