Hamel Husain teaches engineers and product managers how to evaluate AI systems and connect evaluation results to customer needs. His work includes courses and a book about AI evaluations.
Hamel Husain is most useful for teaching engineers and product managers to evaluate AI systems and connect evals to customer needs.
Best for
A B2B AI tooling company upskilling product and engineering teams on evals and AI product discovery.
Not for
Consumer lifestyle app seeking paid social promotion for short video series to broad demographic.
Key takeaways
Public OCR benchmarks can expect absent information and penalize correctly preserved line breaks.
Ask customers to describe a specific past problem, then let evals test the fix.
Question whether an eval dashboard is the best way to share results.
Check first: Confirm whether bespoke workshops or advisory retainers are available, since only course seats and book are documented.
Main ideas
A benchmark score is a claim to audit, not a ranking to trust · Be skeptical of eval dashboards · Ask about a real past problem, then let evals test the fix
Audience
34,579 LinkedIn followers · over 4,500 Students taught in the AI evals course (creator's own figure, co-taught with Shreya Shankar)
Hamel Husain teaches engineers and product managers to evaluate AI systems and connect evaluation results to real customer problems.
His work spans a course and the book Evals for AI Engineers, alongside practical writing on benchmarks, product discovery and how teams can tell whether an AI product is improving.
Read his work for concrete ways to make evaluation more useful than a leaderboard score or a dashboard.
Hamel makes evaluation failures tangible.
In a post on problems with public OCR benchmarks, he highlights tests that expect information absent from a document and scoring rules that penalize a correctly preserved line break.
The lesson is that an overall benchmark score can obscure whether a model performs well on the capabilities a team actually needs.
He also connects evaluation to the decisions that come before building.
Those interviews help teams identify what to solve; evaluations can then test whether the product addresses it.
His advice is practical and sometimes counterintuitive: a post about evaluation dashboards urges teams to question whether a dashboard is the best way to share results. Alongside his book announcement, these examples show a consistent focus: make evaluation serve better product decisions, rather than treating scores as the goal.
Hamel Husain's recurring arguments are about measurement and evidence: audit an evaluation before trusting its number, connect evaluation work to what customers actually said, and teach the parts of the subject that change quickly in live settings while writing down the parts that hold.
Read more
His posts frequently summarize and credit other practitioners, so their methods appear here as his framing of their work rather than his own.
A benchmark score is a claim to audit, not a ranking to trust
Hamel Husain argues that public benchmark scores are weak grounds for choosing a model, because the benchmarks themselves contain errors that distort the ranking.
Read more
His reasoning is that a single overall score hides the capability differences a buyer actually cares about, and that scoring rules and postprocessing can move results more than real model quality does. He makes the case through a talk by Vik Paruchuri, presenting it as an argument he endorses rather than his own finding.
From that talk: one benchmark expects 310 gigawatts of solar capacity when the source document says 303, and 73% of its chart questions ask for values that are not printed on the chart.
A scorer requires "Parish Priest:" to be bold when the words are not bold in the source.
A model ranks last overall on one benchmark while beating the top-ranked model on content and tables.
A 700-line postprocessor adds one to two points, more than the gap between many models, without any improvement in the underlying model.
Hamel Husain treats the dashboard as the wrong unit of analysis.
Read more
Asked what a good eval dashboard looks like for stakeholders, he answers that teams should be very skeptical of dashboards and that with AI available they can do better. The argument follows from his benchmark criticism: a compiled score loses the specifics that would tell you whether the system is improving where it matters.
His answer is a two-sentence push back on the premise of the question, with links to further reading rather than a chart layout.
Ask about a real past problem, then let evals test the fix
Hamel Husain argues that most customer conversations engineers run produce weak signal, because asking what someone thinks of an idea or showing a demo invites polite reaction rather than evidence.
Read more
He describes Teresa Torres's method, attributing it to her: ask customers to describe a specific time they experienced a problem, then walk through what they did.
The interviews are synthesized to prioritize work, which he notes is harder than it sounds and takes practice.
Evals then serve as a second feedback source, testing whether the product solves the problem customers described.
The comparison he draws is between "What do you think of my idea?" and a walkthrough of a specific past problem. He also cites Torres's line that she is no longer designing courses to teach discovery but creating software to teach discovery, which was reported in a separate post on her reimagined education work.
Teach fundamentals in a book, refresh the changing parts live
Hamel Husain splits his teaching by how fast the material moves.
Read more
The live course has been the only guided way to learn evals precisely because the format allows the curriculum to be updated as methods emerge; the book covers fundamentals he expects to hold as models improve. Shreya Shankar is named as his co-teacher for the refreshed course.
Announcing the book Evals for AI Engineers alongside a Bay Area meetup, he explains the live-versus-book division directly. The refreshed course AI Evals for Engineers & PMs covers agentic coding and advanced data analysis taught live, adds a session on job interviews that include evals, and is described as having taught over 4,500 students.
Hamel Husain uses a simple repeated count to argue that evaluation work has moved from a niche interest to a routine expectation.
Read more
His method is counting eval mentions in every Hacker News "Who is hiring?" thread since late 2022 and reporting the change as a share. He reads the pattern as companies that shipped an AI product needing someone to measure whether it is getting better and where it needs work, across engineering, infrastructure, and product management roles.
The share was close to zero in early 2023 and 10.7% of roughly 6,000 AI postings in the most recent month he examined.
If you are choosing a model and see a leaderboard score, this post explains the questions to ask first.
Read more
Hamel Husain shares a talk by Vik Paruchuri arguing that public benchmark scores should not be trusted for model selection. The post lists concrete OCR benchmark failures: ParseBench expects 310 gigawatts when a document shows 303, and most chart questions ask for values not printed on the chart; ParseBench requires “Parish Priest:” to be bold when the source is not; a scorer misses headings on separate lines; dots.mocr ranks last even though it beats the top model on content and tables; olmOCR-Bench penalizes a model for correctly preserving a line break and for copying “Republican party” instead of “Republican Party”; a 700-line postprocessor adds one to two points, hiding lack of model improvement; and OmniDocBench penalizes a correct HTML table more than a Markdown table with a wrong number.
For brands planning a campaign
Useful for a workshop on evaluating models by task, not leaderboard position; Hamel Husain can turn these examples into a buyer-side checklist for engineering and product teams.
This is the clearest starting point for what Hamel Husain teaches in a structured way and where to learn eval fundamentals.
Read more
Hamel Husain announces the book Evals for AI Engineers and a Bay Area meetup on Nov 12.
The post says the live course had been the only guided way to learn evals, and the book covers fundamentals expected to remain stable as models improve.
In-person attendees get a signed copy, a short lecture on evals, and a meet and greet.
For brands planning a campaign
Shows a packaged curriculum and public teaching format; a company could propose a private workshop built around the same fundamentals.
Answers why customer interviews are not the same as demo feedback, and how discovery connects to evals.
Read more
Hamel Husain says engineers often get weak signal by asking “What do you think of my idea?” or showing a demo.
Teresa Torres teaches teams to ask customers to describe a specific time they had a problem and then walk through what they did, which gives stronger evidence about what to solve.
She also teaches how to synthesize interviews to prioritize work.
Once a team builds an AI product, evals test whether it solves the problems customers described.
For brands planning a campaign
Supports a workshop that connects product discovery to eval design; useful for a B2B AI tooling company whose PMs and engineers need a shared method.
Explains why evals have moved from niche to hiring requirement, with a simple data point.
Read more
Hamel Husain counted mentions of evals in Hacker News “Who is hiring?” threads since late 2022.
In early 2023 the share was close to zero; last month it was 10.7% of about 6,000 AI postings.
The roles vary across engineering, infrastructure, and product management.
Hamel Husain says these companies shipped an AI product and need someone who can measure whether it is improving and where it needs work.
The post also describes a refreshed “AI Evals for Engineers & PMs” course with Shreya Shankar, live teaching on agentic coding and advanced data analysis, more than 4,500 students from companies including Google, OpenAI, Meta, and Amazon, and a bonus job-interview session; the next cohort starts September 5.
For brands planning a campaign
Gives a market case for upskilling; the employer list and course format are relevant to a company considering a cohort or advisory engagement.
A short, practical warning about replacing eval thinking with a dashboard.
Read more
Asked what a good eval dashboard looks like for stakeholders, Hamel Husain answers: be very skeptical of dashboards, and with AI you can do better than a dashboard. The post links to further reading and a FAQ.
For brands planning a campaign
Useful as a provocation in an executive session: it pushes teams toward targeted evaluation questions rather than a single dashboard, a theme Hamel Husain can expand for stakeholders.
Shows Hamel Husain’s interest in AI-assisted teaching and the people he learns from.
Read more
Hamel Husain describes Teresa Torres as one of the most prolific professional educators in technology, with more than 200,000 students taught customer discovery, and says she has shaped product management.
The post highlights her personal story of reimagining education with AI, including the quote: “I am no longer designing courses to teach discovery.
I am creating software to teach discovery.
My medium has changed.”
The post links her article and points to further articles about her AI coaching assistant.
For brands planning a campaign
Connects Hamel Husain’s teaching practice to AI coaching products; relevant if a campaign is about scaling education beyond live workshops.
Evaluations as a measurement discipline, not a dashboard score
Hamel Husain's posts argue that an eval number should be audited before it drives a decision.
Read more
He walks through a talk by Vik Paruchuri that found concrete failures in popular OCR benchmarks: one scorer expects 310 gigawatts when the source document shows 303, and 73% of its chart questions ask for values that are not printed on the chart.
Others penalize a model for correctly copying capitalization or preserving a line break, or require a word to be bold when it is not bold in the source.
On one benchmark, a model ranks last overall even though it beats the top-ranked model on content and tables, and a 700-line postprocessor adds one to two points without any improvement in the underlying model.
Husain applies the same skepticism to stakeholder reporting, answering a question about eval dashboards by saying teams should be skeptical of dashboards and can do better with AI available.
Close readings of named benchmark cases with specific numbers, and a question-and-answer format that turns a practice problem into a short recommendation.
For brands planning a campaign
A workshop built on this material should start with an audit of a client's existing eval suite, looking for mislabeled tests, scoring rules that punish correct behavior, and postprocessing that flatters results. That is immediately useful to a company that already has eval tooling but is unsure what its scores mean.
Hamel Husain connects evaluation work to customer research.
Read more
He describes Teresa Torres's method: asking customers to describe a specific time they experienced a problem and walking through what they did, rather than asking what they think of an idea or showing a demo.
The interview results are then synthesized to prioritize work, which he notes is harder than it sounds and takes practice.
Evals come next, testing whether the product solves the problems customers described, and the results are used to close the gap quickly.
He also highlights Torres's account of redesigning education with AI, including the line that Torres is no longer designing courses to teach discovery but creating software to teach discovery.
Summarizes another practitioner's teaching process step by step, then explains how evals fit into that process as a feedback loop.
For brands planning a campaign
For a brief about upskilling product and engineering teams, this supports a workshop that pairs discovery interviews with eval design, so teams learn to derive tests from real customer evidence instead of from internal assumptions.
Live cohort courses and a book as parallel teaching formats
Hamel Husain teaches through live instruction and has written a book to carry the durable parts.
Read more
He announced the book Evals for AI Engineers with a Bay Area meetup, explaining that the live course had until then been the only guided way to learn evals because the live format lets the material be refreshed as methods emerge, while the book covers fundamentals he expects to hold as models improve.
In another post he notes a refreshed curriculum for the course AI Evals for Engineers and PMs, taught live with Shreya Shankar, covering agentic coding and advanced data analysis, with a bonus session on job interviews that include evals.
He reports that the two of them have taught over 4,500 students from companies including Google, OpenAI, Meta, and Amazon.
Live cohort teaching with a refresh cycle, plus a book for the parts of the subject that change slowly.
For brands planning a campaign
A company can use the book as a low-cost starting point for many engineers and bring Husain in live for the parts that need discussion, such as reviewing the team's own eval data or current agentic workflows. This fits a workshop or advisory engagement better than self-paced content alone.
Hamel Husain uses job-posting data to show where the field is going.
Read more
He counted how many AI job postings mention evals in every Hacker News "Who is hiring?" thread since late 2022: close to zero in early 2023, and 10.7% of roughly 6,000 AI postings in the most recent month he examined. The titles vary, spanning engineering and infrastructure roles as well as product management, and he reads the pattern as companies that shipped an AI product needing someone to measure whether it is getting better and where it needs work.
Simple repeated count across a long-running public hiring thread, reported as a percentage change over time.
For brands planning a campaign
Useful as an internal argument for budget: the hiring data shows eval skills are becoming a standard expectation, so training a product or engineering team is a response to a market shift rather than a niche interest.
Documented brand relationships for Hamel Husain are limited and mostly tied to his own teaching work.
Read more
He co-teaches an evals course and received event support from Raindrop for his book launch. The other posts in the reviewed set discuss benchmarks, models, and other people's work without stating a commercial relationship.
AI Evals for Engineers & PMs
Owned business
Hamel Husain calls this "our course" and says Shreya Shankar and he have taught over 4,500 students in it.
Read more
The refreshed curriculum covers agentic coding and advanced data analysis, taught live, with a bonus interview-prep session. The post presents it as a course Husain co-teaches rather than a third-party product.
Hamel Husain's work rewards a reader who wants to evaluate AI systems rather than admire them.
Read more
Start with the posts that walk through broken benchmarks and argue that a single score hides the capability differences that matter; then read the ones connecting evaluation to customer discovery, where the point is to test whether a product solves a problem someone actually described.
Two teaching strands run alongside the arguments: live cohort courses that can be updated as methods change, and a book for the fundamentals that hold as models improve.
That split is itself a useful signal about what to expect — the moving parts get taught live, the stable parts are written down.
Read the hiring-data post last, as context for why evaluation skills have become a routine job requirement instead of a niche interest.
Both treat evaluation as an ongoing practice rather than a launch checkbox.
Read more
Sobhani describes pairing automated evals with human review so quality is checked every time the product changes, and adds bias testing before users see the product — the same insistence Husain makes that a number be audited and tied to the capability a buyer cares about.
Sobhani writes as a product builder inside K-12 education, working from her own team's architecture and research process; Husain teaches evaluation to engineers and product managers broadly and spends more time dissecting where published benchmarks go wrong.
Both are in the business of getting AI skills into the hands of learners at scale, and both describe programs in terms of what students actually built rather than what was taught.
Read more
Eldridge's account of community college teams shipping working apps from an embedded Google AI curriculum is the education-access version of the argument Husain makes from hiring data, that practical evaluation and AI skills have become a baseline expectation.
Eldridge focuses on workforce development and opportunity gaps across California community colleges, with institutional partnerships as the mechanism; Husain focuses on the technical craft of evaluation for people already working on AI products.
Both write about what technology can and cannot replace in learning.
Read more
Sailor's point that tools do not match the human elements of teaching — empathy, understanding, connection — sits beside Husain's caution that an eval score, a dashboard, or a benchmark ranking is not the same as evidence that something works for the people it is meant to serve.
Sailor teaches special education online and writes from day-to-day classroom practice; Husain writes for technical teams about measurement, discovery, and course design at a much larger student count.
The supplied evidence supports one measured audience figure: 34,579 LinkedIn followers, observed 8 October 2026.
Read more
That establishes scale on the platform where this creator publishes; it does not say who those followers are.
Everything else available describes content appeal rather than audience composition.
The posts are written for engineers and product managers who build and evaluate AI products, and the creator's own numbers describe a live evals course, taught with Shreya Shankar, that has served more than 4,500 students, with Google, OpenAI, Meta and Amazon named among their employers.
That is a claim about course participation, not a measured breakdown of the social audience, and should be read as the former.
No geography, age, gender, employer, occupation or seniority data appears in the supplied material, and the follower count cannot support conclusions about any of those attributes.
For a brief that hinges on the audience's professional profile, the practical next step is to ask whether the creator holds recent audience survey results or another documented source covering geography, age, gender, employer, occupation and seniority, and to request recent results for comparable topics, formats and depth before treating this audience as a match.
Hamel Husain's posts describe two paid commercial products: a live cohort course, "AI Evals for Engineers & PMs," co-taught with Shreya Shankar, and a book, "Evals for AI Engineers."
Read more
Both teach teams how to evaluate AI systems.
The posts report that the course has taught more than 4,500 students, names past corporate participants such as Google, OpenAI, Meta, and Amazon, and advertises a next cohort start date; a separate post marks the book launch and a Bay Area meetup.
These enrollment and event links are the only commercial routes named in the material reviewed.
None of the posts states a sponsorship policy, says whether Hamel Husain accepts brand partnerships, lists rates, deliverables, or availability, or points to a partnership contact.
The practical consequence for a buyer: the documented way to work with Hamel Husain is to buy seats in a scheduled cohort or buy the book, and there is no published channel for a bespoke workshop, advisory retainer, or sponsored content.
A B2B AI tooling company wants to hire a technical educator and LLM evals expert to upskill its product and engineering teams, with a workshop or advisory engagement on evals and AI product discovery.
Reasons
The teaching record is documented, not implied. Hamel Husain announced a book, Evals for AI Engineers, and describes a live cohort course, AI Evals for Engineers & PMs, co-taught with Shreya Shankar, reporting more than 4,500 students from companies including Google, OpenAI, Meta and Amazon. A course and a book are the two formats the reviewed posts describe as paid products, and both teach the exact subject this brief asks for.
The eval expertise is concrete. The reviewed posts walk through named benchmark failures with specific numbers: a scorer expecting 310 gigawatts where the source document shows 303, most chart questions asking for values not printed on the chart, a benchmark requiring a phrase to be bold when the source is not, a model ranked last overall despite beating the top-ranked model on content and tables, and a 700-line postprocessor adding one to two points without any improvement in the underlying model. This is the material an evals workshop needs.
The discovery half of the brief is also covered. Hamel Husain describes the interview method associated with Teresa Torres — asking customers to describe a specific time they experienced a problem and walking through what they did, rather than asking what they think of an idea or showing a demo — and explains how the synthesised interviews feed into evals that test whether the product solves the problems customers described.
There is a usable market argument for internal budget. Hamel Husain counted eval mentions in Hacker News hiring threads and reports the share moved from close to zero in early 2023 to 10.7% of roughly 6,000 AI postings in the most recent month examined, across engineering, infrastructure and product management roles. A training request can be framed as a response to a hiring shift rather than a niche interest.
The stance on reporting suits a working session. Asked what a good eval dashboard looks like for stakeholders, the answer given is to be skeptical of dashboards and to do better with AI available. That is a productive starting provocation for a team that already has eval tooling and is unsure what its numbers mean.
Constraints
The documented commercial routes are seats in a scheduled cohort and the book. No post reviewed states a partnership policy, rates, deliverables, availability, or a contact route for a bespoke workshop or advisory retainer, and none points to a partnership enquiry address. Treat a private engagement as unconfirmed and ask directly rather than assuming it is available.
No audience survey data was supplied. The only measured audience figure is a LinkedIn follower count of 34,579 observed 8 October 2026. Nothing in the reviewed material covers geography, age, gender, employer, occupation or seniority, so the brief should not be sold on an audience-composition claim.
When to choose a specialist
Poor fit on the evidence reviewed.
Read more
Brief
A consumer lifestyle app wants a paid social promoter for a short-video series about everyday uses of AI, aimed at a broad demographic audience in a defined age and region, with product placement woven into personal-story content.
Reasons
No sponsorship format is documented. The reviewed posts describe a course, a book and a launch meetup. They do not state a sponsorship policy, a rate card, deliverables, exclusivity terms or a partnership contact, so there is no published basis for commissioning sponsored short-form content.
The audience cannot be matched to the brief. The only supplied audience measurement is 34,579 LinkedIn followers observed 8 October 2026, with no geography, age, gender, occupation or seniority data anywhere in the reviewed material. A brief built on targeting a consumer demographic cannot be assessed, let alone confirmed, from what is available.
The content format is a mismatch. The reviewed material is written technical posts and live teaching about evaluating AI systems. Nothing in it indicates short-form video production, consumer lifestyle subject matter, or personal-story content of the kind this brief needs.
The editorial stance works against the ask. Hamel Husain's published position is sceptical of surface-level signals: audit an eval number before it drives a decision, and be wary of dashboards as a substitute for real measurement. Asking for promotional claims about a consumer AI feature runs against the kind of close, critical reading the reviewed posts model.
External links are unresolved. Several posts point to a talk, an article, a course page and an FAQ. None of those bodies was supplied, so any conclusion about formats beyond the reviewed posts would be unsupported.
A worked campaign example
Campaign angle
Eval readiness audit plus a discovery-to-eval workshop for the B2B AI tooling company, built as a two-part engagement with a low-cost entry point.
Steps
Entry: buy seats in the next live cohort of AI Evals for Engineers & PMs and copies of Evals for AI Engineers. Both are documented paid routes and both give the product and engineering teams a shared baseline before any custom work.
Part one, audit: run the client's existing eval suite against the failure modes described in the reviewed posts. Look for expectations that do not match the source data, scoring rules that penalise correct behaviour such as preserving a line break or copying capitalisation accurately, and postprocessing layers that lift scores without any change in the underlying model. Output is a rewritten set of eval items plus a short note on which existing scores should not be trusted for model selection.
Part two, discovery-to-eval: teach the interview method described in the posts — ask a customer to describe a specific time they hit the problem, walk through what they actually did, then synthesise across interviews to prioritise. Convert the strongest problem descriptions into eval items that test whether the shipped product solves them.
Internal justification: present the hiring data, close to zero eval mentions in early 2023 rising to 10.7% of roughly 6,000 AI postings in the most recent month examined, as evidence that eval skills are becoming a standard expectation for the roles the client is hiring.
Market comparison: a general LLM bootcamp or prompt-engineering trainer will cover more ground, but the reviewed evidence shows no comparable depth on benchmark auditing or on linking product discovery to eval design. Choose the broader trainer if the goal is general fluency; choose this angle if the goal is fixing measurement the client already ships with.
Next step: approach through the cohort and the book first, since those are the routes the material documents, and raise a private workshop or advisory separately. That route is not published, so it needs explicit confirmation before it appears in a plan.
Sabina Sobhani is the closer substitute when the client wants someone who has shipped an AI product in a high-stakes setting and can teach the design decisions behind it, rather than benchmark methodology.
Read more
The reviewed post describes a framework that combines automated evals with human review on every change, explicit bias testing across demographic contexts before users see the product, and context engineering as at least as important as prompt engineering.
That maps onto a team building an AI feature whose output drives a person's next step.
It is also the better pick when the training should be tied to how practitioners actually use the tool: the reviewed material links product quality to how educators support individual students, which is a concrete account of designing for trust under real consequences.
The public teaching record is far thinner than the baseline.
There is no book, no cohort course, no reported student numbers and no published hiring data to justify a training budget.
The section rests on two posts — one on AI product design, one on an education technology award — so there is no basis for judging how this person teaches across a multi-day engagement, whether a workshop format exists, or how the material holds up in front of an engineering audience.
Choose Sabina Sobhani for demonstrated product-building judgement in a regulated, human-facing domain; choose the baseline if the deliverable is a tested, repeatable evals curriculum.
Amanda Eldridge is the better choice when the brief moves away from upskilling an internal engineering team on evals and toward building a supply of AI-skilled people, or toward a community college partnership.
Read more
The reviewed posts describe a programme run across 85-plus California community colleges, with more than 4,000 applicants, nine student teams pitching apps built on Google Gemini and films made with Veo 3 to Google executives, and named student projects such as a pet-symptom checker and an accountability workout app.
That is a detailed, end-to-end account of running a workforce programme with an external partner.
A client whose goal is recruitment, apprenticeship or regional AI talent supply would weigh that delivery record above eval methodology.
Nothing in the reviewed posts covers LLM evaluation, benchmark design or teaching engineers to measure model quality.
The subject is programme delivery, student outcomes and partnership mechanics — not measurement practice.
A client that needs its product and engineering staff to audit an eval suite, spot scoring rules that punish correct behaviour, or connect discovery interviews to eval items would get no subject-matter depth here.
Choose Amanda Eldridge for pipeline and partnership work; keep the baseline for measurement teaching.
Colin Sailor is the better choice when the buyer task is teaching practitioners to deliver instruction online to students with diverse needs.
Read more
The reviewed posts describe five years of online teaching, why online formats remove stressors that students face in physical schools, how families report the approach fitting their situation, and a firm position that tools do not replace empathy, understanding and connection — with a note that students look for human solutions at the end of the school year. For a client whose AI product is used in remote or special education settings, an instructor who has taught in that setting for years is a different and useful input to a training or advisory engagement.
There is no AI evaluation content, no benchmark or measurement material, and no evidence of teaching engineering or product teams.
The machine-learning depth is a single statement about integrating technology and AI into a classroom; the rest is teaching practice and the limits of tools.
It does not answer the original brief about upskilling a product and engineering team on evals, and the reviewed material shows no commercial offer, format or enquiry route that a client could act on.
Choose Colin Sailor for classroom pedagogy in remote and inclusive settings; keep the baseline for evals and AI product discovery.
This page draws on six LinkedIn posts from Hamel Husain's account, read in full from the supplied text, covering a talk on public OCR benchmark failures, an answer on eval dashboards, customer discovery linked to evals, the launch of the book Evals for AI Engineers, a refreshed cohort course, and a count of eval mentions in Hacker News hiring threads.
No publication dates were supplied, so sequence comes from what the posts describe, such as a live-only course followed by a book and a refreshed curriculum.
Every post is treated as this person's own writing; where a post summarizes another practitioner's work, that method is reported as Hamel Husain presents it.
The talk, article, course page and FAQ the posts link to were not available, so nothing here claims to have read them.
Identity, ideas, relationships, similar-creator and campaign-fit sections were used as accepted dependencies; no engagement, rate, audience breakdown or sponsorship policy is described beyond what the reviewed posts state.
Recent engagement
Across their last 3 LinkedIn posts, the median post got 45 reactions and 24 comments. Checked 2026-11-02.
