Video summary
ML Engineer сьогодні: задачі, технології та відповідальність — Олександр Корнієнко, випускник (2024)
Main summary
Key takeaways
Main ideas, concepts, and lessons
1) Purpose and structure of the talk
The speaker shares a personal career story to answer:
- How trajectories in ML/engineering change over time
- What different ML projects have in common
- What employers and candidates should pay attention to
The talk is divided into three thematic blocks:
- Personal ML engineering trajectory + projects
- Miltech (military technology): what it is, what roles/skills matter, and opportunities
- Superhuman (formerly Grammarly): what the ML Engineer role looks like, what the challenges are, and what to learn/aim for
2) How the speaker’s background shaped their ML path
- Initial education: technical information protection (security).
- Key connection: security work (e.g., analyzing data access and preventing unauthorized use) connects logically to later work where NLP and ML analyze language/information.
- The speaker describes a progression of seemingly “different” domains that are actually grounded in similar fundamentals:
- signal processing
- math
- feature engineering
- modeling
3) Project history: “cool projects” across ML domains (common themes)
The speaker lists a chain of commercial projects, emphasizing shared foundations (math, signals, feature representations, pipelines):
- Video analytics / person tracking
- Music recognition / recommendation systems
- Detection suit (“blows”)
- Uses signal processing + spectral analysis and feature selection
- Perception system for processing patient data
- Described broadly as perception + processing pipelines
- NLP / language models / agents (current direction)
Early inspiration
- Learned about OpenCV and noticed similarity between “HAR” concepts and classic detector pipelines.
- Early detector systems were built using:
- cascades of feature functions
- ensembles
- analogy: a CNN as an ensemble element, where a shallow network feeds into ensemble structure
First job hiring story
- A recruiter/interviewer recognized excitement about OpenCV and the similarity to known concepts.
- The initial task involved an ML/POC system for:
- searching for people
- counting people
- e-commerce analytics (“who comes, what products they’re interested in”)
4) Scale and “product thinking” in NLP at Grammarly/Superhuman
In the speaker’s first NLP role, they clarify that ML Engineering is not just model training.
Responsibilities
- Help with data collection and annotation
- Assess annotation quality (agreement/disagreement between annotators)
- Train models
- Deploy to production systems handling ~200,000 requests/sec
Early system approach
- Rewriting text in a desired tone/style using GPT-2
Quality and evaluation challenge
- Defining “style/tone” metrics is subjective.
- They mention an “LM-as-a-judge” idea:
- other LMs assess:
- author knowledge / naturalness
- task validity (does it solve the goal?)
- other LMs assess:
Key pain point
- Maintaining GPT-2 at high traffic is expensive.
Solution direction
- Reduce cost via distillation and smaller specialized models.
Core lesson
- ML engineers in product companies need a full-cycle, ownership mindset:
- understand the problem
- ship
- diagnose failures
- predict failure modes from experience
- Also: learn why users dislike changes—not only when outputs are “technically correct.”
5) War-time/volunteer technology work: using ML for urgent real-world problems
Around Feb 24, the speaker describes a volunteer turning point and actions taken.
Initial efforts
- Reconnaissance concepts using drones to locate threats (mines).
- Attempts at deepfake detection
- They deployed a classifier quickly after the Zelenskyy “surrender” video event.
- Concluded it wasn’t useful—nobody needed it (in that form/time).
Shift: war damage assessment
- Using satellite images to:
- digitize damaged cities (Mariupol mentioned first)
- align pixels into geospatial coordinates
- build pipelines for large image processing
Long-term idea (~since 2017)
- Demining: detect mines from UAVs
- drone sensors (cameras + magnetic sensors)
- mark mines on maps
Explicit lesson
- Value and motivation come from working on problems with no ready-made answers.
- Certain companies/products provided a framework:
- First check whether someone already solved it (open-source/existing code).
- Take responsibility for the result—not just for “doing tasks.”
- “Execution” mindset: don’t wait for step-by-step instructions—own decisions and outcomes.
6) “UCU” / community and fundamentals: how they improved the speaker’s approach
The speaker explains why they later pursued UCU-related education.
- Belief: knowing “how” vs understanding “why” matters.
- They wanted to strengthen foundational gaps, especially mathematics.
- UCU is described as providing:
- a strong community of capable peers/teachers
- pillars:
- linear algebra (practical)
- ethics (including multi-context understanding)
- “computerized corporate finance” framed through opportunity cost and compound interest as investing in knowledge/health/projects
7) Building ML systems at Miltech: what makes it different
The speaker emphasizes Miltech is fundamentally different from standard commercial ML.
Core characterization
- Not “weapon/romance,” but engineering under extreme constraints.
Main enemies
- time and error
- in Miltech, saving quality has a cost measured in lives/health
System expectations
- reliably and quickly
- under uncertainty
- “right the first time”
- no “turn it on later,” no “turn off and forget”
Miltech is more than an algorithm
- Hardware + application methodology
- algorithm doesn’t matter if devices fail in cold/heat/discharge constraints
- User reality
- defenders may use devices with gloves, in darkness, under fire—often with minimal instructions
- usability should be “one button” (conceptually)
- Field testing from day one
- overheating, antenna failures, battery behavior, and other practical issues
- anticipating these at design time is highly valued
Example constraints
- Thick connectors because devices may be dropped into swamp/snow; must remain functional after cleanup in cold.
Data problem and “synthetics” (simulation)
- Miltech-relevant data is extremely valuable.
- Danger: relying on open general datasets (e.g., speech recognition datasets) can cause hallucinations/model degradation.
- Therefore:
- collect field-relevant data
- but when data is scarce, use synthetics/simulation to cover edge cases
Example
- Camera view geometry / small targets
- if a tiny target occupies only a few pixels, a model may miss it
- simulation generates more visible training scenarios for rare events like mines
Quick Miltech market notes
- Roles distribution mentioned: roughly a “third” are software engineers
- Pay: more competitive than earlier volunteer-stage work
- Growth expectation: increasing demand for additional features
8) Technologies Miltech recruiters look for (as stated)
The speaker lists key technology areas:
- Computer vision
- object detection
- feature tracking
- deployment to single-board devices (e.g., Raspberry Pi) with adequate FPS
- Protection
- code suppression / encrypting model weights to prevent reuse by enemies
- ROS (Robot Operating System)
- build robots using consumer/toy platforms to learn:
- data connectivity problems
- stability
- combining data + decision algorithms
- machine/robot control aspects
- build robots using consumer/toy platforms to learn:
9) Superhuman: the ML Engineer role and agent-focused engineering
The speaker explains Superhuman’s mission and their ML work.
Company/product context
- Superhuman is a multi-product company for AI assistants in communication/knowledge.
- Products mentioned:
- Koda
- Grammarly
- Superhuman Mail
Speaker’s role
- in-house LMs
- text improvement and generation
- recommendation systems for production agents operating across millions of emails daily
Key implication
- model changes directly affect how millions communicate
Why not “just connect GPT” for proofreading
- Cost at scale
- tens of millions daily users; even sentence-by-sentence checking becomes extremely expensive
- Latency
- GPT-style services may take seconds vs ~milliseconds for specialized pipelines
- Meaning change risk
- models can alter meaning (e.g., “insist” vs “leave”)
Approach to quality
- Tune a large model into a smaller one:
- teach it the specific task
- avoid changing unnecessary text
- Evaluate with task-specific metrics
- mentions “F05” / precision-like emphasis (precision reportedly more important than recall)
- If needed:
- ensembles of specialized models can improve quality, but increase cost
How agents work (described)
- Agents run multi-step loops:
- receive task
- plan steps
- call tools
- check result
- adjust
- repeat before completing
- Tradeoff:
- more autonomy, but often slower and more expensive than direct models
- especially costly for long email/document checking
10) Engineering challenges when building/using agents
The speaker enumerates engineering problems before and when building agents:
- Non-determinism
- LLMs can fail in unpredictable steps (especially if temperature > 0)
- Quality assessment requirement
- because outputs aren’t reliably predictable:
- prefer deterministic solutions where possible
- implement strong quality control
- use:
- human annotation
- LLM-as-judge evaluation (“language as judge”)
- because outputs aren’t reliably predictable:
- Versioning and model replacement risk
- backend model swaps can unexpectedly reduce quality
- Security and control
- “responsible AI” emphasis: models shouldn’t generate risky/harmful content
11) What the speaker says MLEs do day-to-day
Workstreams described:
- working with data
- building models/classifiers
- building LM agent systems
- deployment engineering
Systems integration and operations:
- choose architecture/backends based on cost and latency
- integrate into products (multi-microservice design, protocols, load balancing)
Operational reality:
- incidents can occur at ~3am; engineers must diagnose and fix quickly.
12) What agents changed in the speaker’s working style
- “Code review” habit
- agents can make changes that are hard to control
- code review assistants help prevent breaking changes
- Codebase analysis
- hidden documentation and unknown stacks require asking people
- agent assistants act as quick guides to integrate safely
- Faster routine development
- previously: days to build boilerplate for tricky integration/breaking-model issues
- now: near minutes to half an hour to connect services and test
- LLM tooling for research engineering
- logs, experiment tracking, analysis
- assistant can help train/improve models from model + data
- “downtime” automation idea: wake up and the job is done
13) What employers look for (explicit interview guidance)
The speaker provides structured recruiting signals.
What to demonstrate
- Values fit
- not “cultural vibe,” but mission understanding
- Technical thinking
- solve from the task statement, not just know tools
- be willing to inspect references/approaches (e.g., Face/Hub-style resources)
- Impact
- explain what you did and how you measured outcomes
- Responsibility and explainability
- justify decisions
- explain alternatives and why you chose your approach
Red flags / negatives
- Cannot explain why decisions were made
- Proposes technology without understanding it
- Doesn’t describe teamwork (focuses only on self)
- Cannot analyze mistakes or reflect on what went wrong and what was learned
How to prepare (practical suggestions)
- Build pedagogical projects (achievable, routine problems) and mention them in interviews
- Use:
- volunteering
- internships
- open problems as course/diploma projects
- Learn from people around you (company culture and mentorship matter)
- Even without commercial experience:
- show projects
- measurable outcomes
- reflection and learning
Methodologies / instruction-like lists (as presented)
Miltech “how to think” (implied checklist + steps)
- Treat Miltech as engineering under extreme consequence
- Anticipate constraints early:
- environment (cold/heat)
- power/battery behavior
- physical failure modes (connectors, cables, field damage)
- Iterate with field testing from day one
- Collect relevant data:
- rely on field-specific data (avoid generic datasets that degrade quality)
- When data is insufficient:
- use synthetics/simulation to generate coverage of rare/small targets
- Ensure usability for real users:
- minimal training
- robust interaction assumptions (gloves, darkness, stress)
Career/employment preparation steps (explicit-ish)
- Build at least one:
- pedagogical project (practical, solvable)
- volunteer work or internship experience
- team-based projects (course/diploma) showing collaboration
- In interviews, be ready to provide:
- the problem
- your decisions
- alternative reasoning
- metrics/impact
- lessons learned
“Agents vs direct models” decision framing
- If you can solve with targeted tuned models:
- do it (cheaper, faster, more controllable)
- If tasks require multi-step tool use and autonomous loops:
- use agents, but plan for:
- non-determinism
- quality assessment gates
- security controls
- versioning risks
- use agents, but plan for:
Speakers / sources featured
- Primary speaker: Олександр Корнієнко (Alexander Korniienko) — ML engineer / lecturer
- Mentioned person (brief): Руслан (referenced in the opening anecdote)
- Mentioned people (brief):
- Олександр “Andrey Kurorychkin” (invited the speaker to teach/lecture)
- Ярік Гермило (Yarik Hermilo) (credited for a talk about code review/control)
- Назарій Друщук (Nazariy Drushchuk) (responsible AI)
- Companies/products mentioned as sources/contexts:
- OpenAI (GPT-2 mentioned; GPT-4 mentioned)
- OpenCV
- Grammarly (spelled in subtitles “Gramerly/Gramrly,” later referred to as rebranded “Superhuman”)
- Superhuman
- UCU (education/community; exact full org name unclear in subtitles)
- ROS (Robot Operating System)
- Raspberry Pi
- Claude (mentioned as an assistant in the agent/tooling discussion)
- Face/Hugging Face-style resource (referenced as “hface” in subtitles; likely Hugging Face)