Команды AI-агентов для роста метрик и карьеры! AI Director @ Deliveroo

SUMMARY

Dima Zborovsky, AI Director at Deliveroo (now DoorDash), shares insider strategies for building AI agent teams in Cursor to boost business metrics, simulate executive decisions, analyze sales, and advance careers through practical demos and a upcoming challenge.

STATEMENTS

Dima Zborovsky is an engineer deeply involved in AI throughout his career, working at Intel, Yandex, a startup sold to Bermarket, and now leading AI teams at Deliveroo across ten countries, acquired by DoorDash.
Deliveroo operates in ten countries with a focus on Data Science and ML teams, where Dima oversees machine learning initiatives.
Dima maintains a vast network of engineers from top AI labs like AI Lab, XII, and DeepMind, providing insider perspectives on what truly works in corporations.
Many content creators share AI overviews and non-professional use cases, but Dima offers an "insider look" based on real corporate and lab experiences to drive mass adoption.
At the end of the session, Dima promises insights on starting the new year with deep AI immersion and details a small challenge with the community.
Cursor is a text editor similar to notes apps but enhanced with AI integration for interacting with documents.
Dima uses Cursor to manage massive communications as an ML leader, simulating executive reviews before real meetings to prepare for questions and pitch experiments effectively.
In Amazon-like cultures at Deliveroo, detailed documents are required for rolling out experiments company-wide, involving reviews from CEO, CFO, and others.
Dima created a "virtual board of directors" using AI agents modeled on real executives to anticipate questions, spending just 10 minutes preparing, which led to a promotion by appearing always in control.
AI agents in Cursor are simple text files with descriptions of goals, common questions, and context about specific people, building a "second brain."
For an experiment analysis on dynamic ad load increasing revenue by 20 million pounds annually, Dima tags the file and invokes a "round table" logic for agents to discuss in rounds.
The round table setup includes agents carrying conversations, providing actionable advice, and highlighting potential questions, formatted using a Jeff Bezos style guide for executive communication.
Agents simulate discussions: CEO asks strategic questions, CFO focuses on financials, Data Scientist nitpicks methodology, leading to summarized positions, risks, and feedback.
Intersection insights from agents reveal shared concerns, which Dima spends most time on, as they uncover novel ideas from combined perspectives.
Cursor processes transcripts from weekly executive meetings to identify key priorities and relevant contacts, replacing the "blank slate" problem in decision-making.
Dima reads and verifies agent outputs but credits them for eliminating initial preparation hurdles, generating Slack posts and executive summaries.
Agents improve weekly as Dima adds more context, like transcripts, making them smarter over time.
In Western tech (US, UK), AI proficiency is mandatory for budgeting; teams plan higher productivity per headcount in engineering, marketing, HR, and content roles.
Companies like Google DeepMind require live AI-coding in interviews; Uber product leads must AI-code full prototypes solo, with engineers providing feasibility checks.
Shopify interviews involve collaborating with an LLM; Amazon, Netflix demand routine AI integration explanations from all roles, tracking metrics and tools.
AI effects come mostly from basic tools like custom GPTs for routine automation, saving hundreds of developer hours, not complex multi-agent systems.
Deliveroo's gains stem from simple pipelines for routine tasks, rarely needing agents; orientation in tool diversity is key for fit-for-purpose applications.
In HR and marketing, unacceptable levels include not using AI for basic tasks beyond biased resume scanning; adaptation is essential for scaling in large companies.
Dima's intensive built agent systems and automated routines over four weeks; now shifting to an offline challenge for broader community sharing of small tasks.
The challenge aims to build tool awareness, first steps for assignments like parsing news or client feedback, changing patterns like using Cursor for text tasks.
For sales teams with 3-4x performance variance, AI analyzes call transcripts against scripts, generating actionable feedback plans with quotes, scaling what once took thousands of hours.
Feedback collection bots pull from platforms daily/weekly, summarizing positives, negatives, feature requests, and key insights into Telegram channels.
Personal use: Voice notes to a Telegram bot create clean transcripts, summaries, and reading suggestions in Notion for idea capture and post writing.
The community challenge includes two live sessions, eight offline tasks on second brain, sales analysis, RAG chatbots, Cursor workflows, competitor scraping, feedback automation, viral content pipelines, and building a startup MVP.
Early bird pricing until Dec 31; formats include general community access and limited personal sessions; bonuses for referrals.

IDEAS

Simulating a virtual board of directors with AI agents anticipates executive scrutiny, turning preparation into a 10-minute ritual that boosts promotion chances.
Agents evolve into a "second brain" by accumulating personal context, transcripts, and experiences, offloading cognitive overload from human memory limits.
Round table discussions among AI personas reveal intersection insights, like combined CFO-CEO concerns, sparking ideas humans might overlook in isolation.
Basic custom GPTs yield massive routine savings—hundreds of developer hours—outpacing complex agents in real corporate wins.
AI proficiency shifts hiring from "nice-to-have" to budgeted expectation, demanding proof why AI can't replace a new hire's role.
Product managers now AI-code full prototypes solo, with engineers as side consultants, blurring non-technical role boundaries.
In sales, a single prompt analyzes transcripts for script adherence, generating personalized feedback that scales coaching beyond manual limits.
Cursor as an "operating system" accesses all files, Notion, emails, and even scrapes competitors, unifying workflows into one interface.
Weekly executive transcript synthesis identifies priority contacts, enabling proactive relationship-building without manual review.
Voice-to-Notion pipelines clean rambles into structured notes with book recommendations, turning walks into productive content ideation.
Community challenges democratize AI by breaking complex builds (like RAG bots) into guided, non-technical steps for non-coders.
Embeddings and autocompletion form the core of text-based AI, explainable in minutes, empowering non-tech pros to design systems.
Feedback bots automate sentiment and feature extraction from platforms, delivering weekly Telegram summaries to inform product decisions instantly.
Viral content agents ingest random thoughts via audio/text, outputting editable posts with generated images, enforcing regular LinkedIn posting.
Building a full startup MVP—GitHub repo, backend logic, frontend, hosting—in two hours via AI collaboration turns ideas into monetizable products rapidly.
Hallucination elimination via round tables and evidence-based rules makes AI outputs reliable for high-stakes executive pitches.
Personal health trackers in Cursor analyze bloodwork, Whoop data, and sleep transcripts for tailored advice, integrating life metrics seamlessly.
Competitor scraping via Cursor prompts generates landing page screenshots and differentiation feedback in one go, accelerating market intel.
AI agents mimic executive styles (e.g., Bezos guide) to draft communications, ensuring polished outputs that impress C-level reviewers.
Offline challenges foster peer support, with live breakdowns turning individual experiments into collective learning loops.
Notion integration via MCP allows Cursor to query hundreds of past notes for pattern analysis, augmenting personal knowledge bases.
Sales performance gaps (3-4x) stem from script deviations, auto-detectable via AI, enabling bottom-up coaching at scale.
Corporate self-hosting of tools like NNN sidesteps cloud restrictions, vital for regions like Russia with access limits.
Early bird communities build vibrancy through nominated extras, evolving curricula based on participant needs.
AI replaces quantity of subordinates with quality of agents, redefining leadership metrics in tech.

INSIGHTS

AI agents as executive simulators not only prepare pitches but cultivate an aura of control, directly correlating to career accelerations like promotions.
Accumulating context turns static tools into dynamic second brains, exponentially increasing decision quality without proportional human effort.
Multi-agent round tables synthesize diverse viewpoints, uncovering emergent insights that single perspectives miss, mirroring real board dynamics.
Routine automation via basic AI trumps sophisticated agents in ROI, emphasizing tool fit over complexity for scalable gains.
Mandatory AI budgeting forces role evolution, where non-technical pros must demonstrate AI augmentation to justify headcount.
Solo AI-prototyping empowers autonomy but highlights infrastructure limits, shifting collaboration from creation to validation.
Transcript-driven sales coaching reveals behavioral gaps at scale, transforming expensive manual reviews into actionable, evidence-based plans.
Cursor's file-access unification eliminates tool-switching friction, positioning it as a personal OS for knowledge workers.
Automated contact prioritization from meeting transcripts proactive-izes networking, turning passive data into strategic opportunities.
Voice ideation pipelines lower content creation barriers, converting ephemeral thoughts into persistent, enriched assets.
Guided challenges bridge theory to practice, building confidence in AI design for non-experts through incremental wins.
Embeddings enable semantic search in vast texts, foundational for RAG systems that make knowledge retrieval intuitive.
Daily feedback aggregation distills user voices into prioritized actions, closing loops between complaints and innovations.
Thought-to-post agents enforce consistency, leveraging AI for volume while preserving human creativity in edits.
Rapid MVP assembly via AI democratizes entrepreneurship, reducing barriers from months to hours for validation.
Evidence rules in agents ensure factual grounding, critical for trust in corporate AI applications.
Integrated health analytics in AI tools personalize wellness, correlating biometrics with habits for optimized living.
One-prompt competitor analysis accelerates benchmarking, providing visual and textual diffs for agile responses.
Style-guided drafting aligns communications with cultural norms, enhancing executive influence.
Peer-nominated evolutions in communities create adaptive learning ecosystems, sustaining long-term engagement.

QUOTES

"I thought: 'I know everything about my top managers, I spend a lot of time with them. I have tons of transcripts to understand how they think... What if I make my virtual board of directors?'"
"Thanks to this, I even got a promotion, they gave me more teams, because people get the impression that everything is always under control."
"They just don't really understand well that I just have a team of AI agents that I communicate with before going to these people and selling something to them."
"Roundtable - this is the big logic where I describe how these agents should help me prepare for a potential conversation."
"In general, I read all this, especially the questions, but now, due to the large amount of time, I want to show you what happens next."
"On 90% the questions that these people discuss here match exactly."
"This replaces my second brain. Understandably, I read all this. Understandably, I check these things somehow, but the problem of starting from a blank slate is completely gone."
"Not AI that will take your job, but people who know how to use artificial intelligence."
"We require this at the financial level. That is, in our budget planning for the same number of people, we lay in greater productivity."
"You must prove to me why AI can't do this before opening a position for a person."
"The bulk of the win in routine, in daily life, will come from ordinary basic tools like custom GPT and the like."
"AI will not make people unnecessary, but those who know how to use AI."
"We implemented a thing that I'm literally going to show you now, in literally 3 minutes."
"This is something that would have taken thousands of hours before, and therefore no one did it."
"Cursor is access through an assistant to all files that are either automatically pulled here or are simply on your computer."
"This is not an operating system, but this is an operating system."
"The main goal of this community is for people to get this breadth of view on different tools."
"We will collect a startup with you. That is, we will connect and in 2 hours people will make a description of the system from scratch."

HABITS

Accumulate transcripts and context into AI agents weekly to evolve them into a smarter second brain for decision-making.
Spend 10 minutes simulating executive round tables before real meetings to anticipate questions and refine pitches.
Tag files in Cursor and invoke round table logic for any incoming experiment or document to generate actionable advice.
Read and verify AI outputs, focusing on intersection insights from multi-agent discussions for novel ideas.
Use Cursor to query past meeting transcripts for top contacts and draft messages, building proactive networking.
Analyze sales call transcripts against scripts via single prompts to create personalized feedback plans for team members.
Automate daily/weekly feedback pulls from platforms into summarized Telegram reports for ongoing product insights.
Record voice notes during walks to a bot for automatic transcription, summarization, and reading suggestions in Notion.
Post regularly on LinkedIn by feeding random thoughts into AI agents that generate editable drafts with images.
Review bloodwork, sleep data, and fitness transcripts in Cursor for personalized health advice and habit correlations.
Nominate and adapt community tasks based on emerging tools or needs to keep learning dynamic.
Start interactions with evidence-based rules in prompts to eliminate hallucinations and ensure factual outputs.
Budget AI tools like Cursor at $18/month for pro features, selecting models via OpenRouter for cost efficiency.
Integrate Notion and other apps via MCP for seamless querying of personal knowledge bases.
Enforce script adherence in sales coaching by auto-generating plans with exact quotes from underperformers.

FACTS

Deliveroo, acquired by DoorDash (valued at $100 billion), operates in ten countries with Dima leading ML across teams.
Dima's network spans engineers from AI Lab, XII, DeepMind, providing corporate-grade AI insights.
An AI-simulated experiment pitch saved Dima hours, leading to a promotion and more teams under his purview.
Agents matched 90% of real executive questions in simulations, despite using demo data.
Western companies budget for higher AI-augmented productivity, reducing headcount needs in engineering, marketing, HR.
Uber product leads now AI-code full use cases solo, a shift from needing engineers or designers.
Shopify requires LLM collaboration in interviews; Google DeepMind tests live AI-coding.
Custom GPTs saved hundreds to thousands of developer hours at Deliveroo on routine tasks.
Sales top performers convert 3-4x better than bottoms; AI transcript analysis revealed script non-adherence as key gap.
Russia's IT gems often skip scaled transcript analysis despite potential for massive efficiency.
Embeddings power 99% of text AI work, alongside autocompletion as foundational mechanics.
Cursor pro costs $18/month, with text token processing at ~0.25 cents per run.
NNN enables self-hosting in restricted regions like Russia, bypassing cloud bans.
A single feedback bot run costs a quarter cent, processing hundreds of entries.
Community intensives sold out in a week, now scaled to broader offline challenges.

REFERENCES

Cursor: Text editor with AI integration for file-based agents and workflows.
Round Table: Custom logic file in Cursor for multi-agent discussions.
Jeff Bezos Style Guide: Document from Amazon investment outlining executive communication formats.
Custom GPTs: Basic OpenAI tools for routine automation in engineering and beyond.
Whoop: Fitness tracker transcripts fed into Cursor for health analysis.
Notion: Integrated via MCP for note querying and voice-to-structured content pipelines.
Telegram Bots: For feedback summarization, voice note processing, and community interactions.
Google Sheets/Docs: Source for daily feedback pulls and experiment snapshots.
OpenRouter: Platform for accessing multiple AI models efficiently.
Claude (Anthropic): Models used in Cursor for superior text handling.
GPT-4.5 / Opus: High-end models for complex agent simulations.
NNN: Self-hosted tool for corporate pipelines in restricted environments.
Confluence: Corporate doc system, integrable via MCP or export to Cursor.
GitHub: Repo setup for MVP building in community challenges.
LinkedIn: Platform for AI-generated viral content posts.
RAG (Retrieval-Augmented Generation): Technique taught for chatbots using embeddings.
MCP (Multi-Cloud Protocol?): Bridge for Cursor to external services like Notion.
WBRs: Weekly Business Reviews, transcripts analyzed for priorities.
Sales Script: Reference document for transcript evaluations.
Base Rules: 250-line file in Cursor for hallucination control.
Community Sprints: Platform hosting the AI agent challenge.
Yandex/Intel/Bermarket: Past employers inspiring Dima's career path.
DeepMind/AI Lab/XII: Networks for engineering insights.
Shopify/Uber/Amazon/Netflix/Google: Market examples of AI hiring practices.

HOW TO APPLY

Install Cursor as your primary text editor and import key documents like transcripts and style guides into a dedicated folder.
Create agent files for key executives: Describe their goals, common questions, and personal context based on past interactions.
For any incoming experiment or report, tag the file in Cursor and prompt: "Use round table to analyze with [agents], provide actionable advice and questions."
Invoke the round table logic to simulate multi-round discussions, focusing on intersections for shared insights.
Review agent outputs for key positions, risks, and questions; forward unresolved ones to your real team for answers.
Use transcripts from meetings: Prompt Cursor to "Go to WBR folder, identify top two contacts for this XP and why, then draft messages."
For sales analysis, upload snapshots and transcripts; prompt: "Identify lowest performer using sales script, create actionable plan with exact quotes."
Set up a Telegram bot to pull feedback from Google Sheets daily: Summarize positives, negatives, features, and send to channel.
Record voice notes to a bot: Instruct it to clean transcripts, summarize ideas, suggest readings, and save to Notion.
For content: Feed thoughts into an agent pipeline to generate post drafts, add images, and schedule LinkedIn shares.
Build a RAG chatbot: Load knowledge base into context, explain embeddings briefly, then iterate to vector store for scalability.
Scrape competitors: Prompt Cursor via MCP to visit sites, screenshot landings, and compare to your own for feedback.
Assemble an MVP: Describe system in Cursor, generate GitHub repo, backend logic, frontend, connect hosting, and add payment stub.
Join the community: Complete offline tasks, share reports in chat, nominate extras, and attend lives for breakdowns.
Optimize costs: Select models like GPT-4.5 via OpenRouter, use caching for repeats, limit to evidence-based prompts.
Integrate health data: Upload analyses and Whoop transcripts to Cursor, query for habit correlations and recommendations.
Enforce rules: Start all prompts with base rules file to ground outputs in evidence and reduce hallucinations.

ONE-SENTENCE TAKEAWAY

Leverage AI agents in Cursor as a second brain to simulate executives, automate routines, and propel career growth in tech-driven firms.

RECOMMENDATIONS

Adopt Cursor immediately as your AI-enhanced OS to unify files, transcripts, and workflows for effortless knowledge access.
Build personalized agent personas from real interactions to preempt challenges in high-stakes pitches and decisions.
Prioritize basic tools like custom GPTs for quick routine wins before scaling to multi-agent systems.
Simulate board meetings routinely to refine communications, aligning with Bezos-style guides for executive impact.
Analyze sales transcripts weekly with script references to scale coaching and close performance gaps.
Automate feedback aggregation from platforms into digestible summaries to inform product priorities proactively.
Use voice bots for ideation capture, turning casual thoughts into structured notes and content pipelines.
Learn embeddings and RAG basics to design scalable chatbots, even as a non-technical user.
Scrape competitors via prompts for ongoing market intel, generating visual diffs to sharpen your edge.
Enforce evidence-based rules in all AI prompts to eliminate hallucinations and build trust in outputs.
Post consistently on LinkedIn using AI-generated drafts to amplify personal branding without burnout.
Integrate personal data like health metrics into AI for holistic habit optimization and life insights.
Join AI communities for guided challenges, sharing experiments to accelerate collective learning.
Self-host tools where needed to navigate access restrictions and maintain data sovereignty.
Prototype MVPs collaboratively via AI to validate ideas rapidly and attract opportunities.
Budget for pro AI access but optimize with model selection to keep costs under $20/month.
Evolve agents iteratively by adding context, ensuring they grow smarter with your expertise.
Nominate and adapt learning paths in groups to stay ahead of emerging tools.
Verify AI outputs manually for critical uses, blending human judgment with machine efficiency.
Start the new year with small AI immersions, like the proposed challenge, for sustained momentum.

MEMO

Dima Zborovsky, an AI engineering veteran who has shaped machine learning at Intel, Yandex, and now DoorDash's Deliveroo across ten countries, views artificial intelligence not as hype but as an indispensable lever for business and personal ascent. In a candid live session, he demystifies AI agents—simple text-based personas in tools like Cursor—as extensions of the mind, simulating executive deliberations to pitch revenue-boosting experiments with precision. What once demanded hours of frantic preparation now unfolds in minutes: agents modeled on real C-suite figures debate dynamic ad loads projecting 20 million pounds in annual gains, surfacing financial risks and strategic queries that mirror boardroom realities.

This "virtual board of directors" isn't science fiction; it's Zborovsky's daily edge, crediting it for a recent promotion by projecting unflappable control. Drawing from Amazon's document-driven culture, he invokes round-table logics where agents converse in rounds, distilling intersections—like shared CFO-CEO concerns—into actionable advice. Hallucinations vanish through evidence-based rules, a 250-line safeguard born of trial and error, ensuring outputs align 90% with actual executive probes. Beyond simulations, Zborovsky's second brain ingests weekly meeting transcripts, pinpointing priority contacts for proactive outreach, erasing the dread of blank-slate starts.

In the cutthroat Western tech landscape—from Uber's solo AI-prototyped use cases to Shopify's LLM-collaborative interviews—AI proficiency is no bonus; it's budgeted necessity. Zborovsky warns that firms like Google DeepMind test live coding, while non-technical roles in HR and marketing face "unacceptable" ratings without basic automations. Yet the real wins stem from simplicity: custom GPTs slashing thousands of developer hours on routines, not elaborate multi-agent setups. At Deliveroo, pipelines parse feedback or sales calls, revealing why top performers convert 3-4 times better—often mere script deviations—scaling coaching that once consumed fortunes.

Zborovsky's demos illuminate Cursor's quiet revolution: an "OS" bridging files, Notion, and external APIs via MCP protocols, querying bloodwork alongside business docs for tailored health insights or competitor scrapes yielding landing-page diffs. A single prompt transforms voice rambles into Notion-ready posts, suggesting reads on cultural shocks like his London Christmas tree heist. For sales teams, transcripts against scripts birth personalized plans with verbatim fixes, a task prohibitive before AI. Feedback bots, runnable for pennies weekly, tally sentiments and feature pleas from platforms, fueling iterations without manual drudgery.

Scaling these for corporations, Zborovsky consults Russian giants—often IT powerhouses skipping such basics—on self-hosted tools like NNN to dodge cloud bans. Embeddings and retrieval-augmented generation, demystified in minutes, underpin chatbots devouring knowledge bases. His mantra: AI doesn't displace jobs but elevates users, redefining leaders by agent arsenals over underlings.

To democratize this, Zborovsky launches an offline challenge: eight tasks building second brains, RAG bots, scraping pipelines, and even a full MVP—from GitHub repo to payment stub—in two hours. Two lives anchor it, with community chats for breakdowns and nominations, early-bird access until December 31 fostering a vibrant ecosystem. Personal sessions offer bespoke guidance, but the ethos is communal—peers dissecting experiments, inviting guests from Revolut or Anthropic.

This isn't abstract futurism; it's Zborovsky's lived acceleration, where AI turns overload into opportunity. As DoorDash's $100 billion shadow looms, he urges immersion: start small, iterate relentlessly, and let agents amplify human ingenuity. In a world blurring coder and creator, adaptation isn't optional—it's the new career currency.