Shirochenko Dmitriy

I completed Meta Back-End Developer! Check out my certificate coursera.org/share/04ba0649… #Coursera

Arxiv article link below 👇 : arxiv.org/html/2508.1618…

In collaborative engineering, mismatched models from suppliers and OEMs create chaos. This LLM framework fixes it with "soft alignment": lightweight mappings layered on top, no originals altered. Preserves everything while unlocking seamless integration. Link in next tweet.

AI's curve is steep: agents and RAG matter, but guardrails and inference optimization make you production-ready. #MLOps #LLM

Datafold's moat stacks three hard-to-replicate edges: - Proprietary SQL compiler does column-level lineage on all queries at scale, building a data dependency knowledge graph. Technically brutal to execute; compounds as more queries feed it. - Deep CI/CD integration (GitHub PRs, dbt) embeds it in dev workflows, driving high switching costs via daily habit lock-in. - Founder Gleb's Lyft creds: Built identical internal tools from lived pain, baking in rare domain expertise. Competitors copying features miss the full flywheel. Operators: Prioritize tools this sticky early. #DataEng

Data quality isn't a standalone monitoring tool. It's a byproduct of solid data engineering workflows. Core insight from Datafold: Most issues stem from code bugs (SQL, Scala, Python), not infra failures. Fix by shifting quality left into CI/CD via data diffing in pull requests, before deployment. This bucks the trend of reactive tools like Great Expectations or Monte Carlo, which alert post-damage. Prevention at source wins. Toughest execution pain: Seamlessly embedding into devs' daily flows so tests actually run, not skipped. Missed opportunity for data teams: Over-investing in detection over workflow integration. #DataEng

Breaking down Datafold (@datafoldcom), a YC S20 AI startup automating data engineering to prevent production bugs. What they do: AI agents handle testing, migrations, and optimization with data diffing (prod vs. dev), column-level lineage, and automated code review. Who it's for: Data and analytics engineers at data-heavy orgs, from startups to enterprises (Perplexity, Disney, Patreon, Thumbtack, Substack). Edge from founders: Gleb Mezhanskiy and Alex Morozov built data platforms at Lyft, Autodesk, Phantom Auto. Key operator takeaway: Manual diffs and lineage tracking waste hours on deploys; this catches issues pre-prod, scaling execution for growing teams. #AI

Arxiv article link below 👇 : arxiv.org/html/2505.1083…

General User Models quietly build a rich portrait of you from screenshots and emails, all processed on-device for total privacy. Spot a wedding invite? It infers suit rentals and suggests them at the right moment. The future of computing: proactive without prying. Article link next tweet.

Skip LLM hype. Builders prioritize agents, RAG, and inference optimization. #LLM #MLOps

[Startup's] moat: Multi-tenant ML models drive network effects. Each brand's catalog and shopper data trains shared models across customers, enabling compound learning single-tenant competitors cannot replicate. Key edges compound this: - Early YC/Tiger funding hired elite ML talent (Oliver joined Klarna at 15, Anton is serial AI founder). - Two-year lead building fashion-specific computer vision, ahead of generic search rivals. - CEO/CTO as AI researchers: Full stack rebuilds keep pace with model advances; traditional SaaS stuck in legacy architectures. - Scandinavia's dense, design-focused fashion brands provided ideal wedge for early traction. - YC 2x Top Company status proves execution speed that scales advantages over time. Operators note: Technical depth avoids the rebuild lock-in pain most SaaS faces in AI shifts. #AI

Depict's core insight flips e-comm rec engines on their head. Traditional ones demand massive clickstream data. Result: cold-start hell that locks out 99% of retailers without years of user history. Depict builds product understanding via deep learning and computer vision on images, descriptions, and attributes. Delivers strong recs and search from day one, no behavioral data required. Tradeoff exposed: For most retailers, what products *are* beats how users clicked them. Inverts the data moat entirely. #AI

Thread: Breaking down Depict.ai (@depictai_status), an AI play on e-commerce search and merchandising for fast-moving brands. What they do: GPT-powered natural language search plus automated visual merchandising. Turns product data and customer behavior into intuitive shopping flows. Replaces fragmented manual tools that break at scale. Who it's for: Fast e-com like fashion (Gina Tricot, Toteme, Jaded LDN, Billionaire Boys Club) and staples (Office Depot, Staples). Founders: Oliver Edholm (CEO, Klarna AI researcher at 15, high school dropout) and Anton Osika (CTO). Operator note: Manual merch tools fail on high-velocity inventory; this automates the pain point end-to-end. #AI

Your favorite AI tool today will be a commodity feature tomorrow. The most durable skill isn't mastering one platform, but adapting quickly to the next. #AI #MLOps

AI boosts coding speed, but builders must master infrastructure, security and scalability. #Security #MLOps

inFeedo's moat stacks data, IP, networks, and stickiness that new AI HR players can't replicate fast. - 9 years of People Science research + 80M employee responses: a compounding data flywheel. Each Amber conversation sharpens attrition predictions; competitors launch from scratch. - Defensible IP: 26 psychologist templates + 7-driver/45-element framework, validated in 3 peer-reviewed papers and Harvard-taught. - Network effects: 330+ CHROs share anonymized benchmarks, boosting prediction accuracy at scale. - Workflow entrenchment: Slack/Teams/HRIS/ATS integrations with 98% auto-resolution for 130K+ employees. 90% response rate (vs 30% industry avg) locks in via better data > predictions > retention > ROI. Operators note: this virtuous cycle turns early adoption into near-impossible switching costs. #AI

Most employee engagement tools equate vocal feedback with engagement. inFeedo flips this: silence is the strongest attrition signal. Non-responders are 3x more likely to quit, yet traditional surveys chase response rates and miss it entirely. They bet employees trust AI (Amber) more than humans or anonymous forms. It's judgment-free and conversational, inverting the idea that bots can't handle sensitive feedback. Engagement also ties to milestones like onboarding, promotions, or manager changes, not annual surveys. Real-time, journey-stage listening beats calendar pings. Missed opportunity for HR tech: optimize for the quiet signal, not just completions. #AI

Breaking down inFeedo (@inFeedo), a strategically important AI play in enterprise HR. They build an employee experience platform for CHROs and people teams to predict attrition, automate support, and track engagement continuously. Flagship product Amber is a conversational AI that runs personalized check-ins in 34 languages, surfacing real-time sentiment, flight-risk signals, and actionable insights. Serves 330+ CHROs at enterprises like Freshworks and Max Fashion across 60+ countries, swapping clunky annual surveys for always-on listening. The edge: scales global teams without the execution pain of manual pulse checks. #AI