About 15 months in the past, I left my full-time job as a machine studying crew lead with the objective of doing unbiased / freelance information science consulting. Since then, I’ve gotten lots of questions on what which means and entails. I’ve not discovered an excessive amount of details about this kind of work, apart from Greg Reda’s incredible submit. I hope this weblog submit solutions a few of these questions for anyone desirous about changing into or hiring an information science marketing consultant.
Full disclosure: I solely actually consulted for the primary ~8 months. I spent the subsequent 6 months of my time beginning and stopping a startup, which would be the topic of a future submit.
What’s freelancing, consulting, contracting, and so on…?
I consider this work as current on a spectrum:
contracting -> company work -> consulting
Sorry to be complicated, however I’ll find yourself utilizing “freelancing” and “consulting” interchangeably all through this submit to explain the work that I did which spanned this spectrum. FWIW, I’d advocate calling your self a marketing consultant as a result of it sounds fancier than the rest.
Contracting
For no matter cause, it’s typically simpler or preferable for firms to rent workers who’re part-time and/or will not be straight employed by the corporate doing the hiring. For instance, that is widespread in Washington DC the place many individuals don’t work straight for the federal authorities however are as an alternative contracted out to the federal government by non-public firms. For me, contracting preparations felt pretty just like being an worker (or fractional worker), albeit considerably faraway from day-to-day politics.
One in all my first freelance gigs consisted of me being a sub-contractor for a Fortune 500 firm. This firm needed some assist forecasting efficiency of their promoting campaigns. On account of… “paperwork”, it was troublesome for this firm to straight rent me as a contractor. As a substitute, they’d an “permitted vendor” firm which employed me after which sub-contracted me out to the shopper. Truly, to be clear, this permitted vendor firm employed my firm, Rosenthal Knowledge, LLC. So many layers of trivia.
On this association, I bought an organization e mail tackle, I attended weekly conferences (nearly, as this firm was on the alternative coast), and I had coworkers that I labored with. As I stated, it felt like being a fractional worker. I billed the permitted vendor for X hours / week till my contract was up or we renewed it.
Company Work
Contracting work might be the least dangerous, and consequently least profitable facet of freelancing. You sometimes signal a contract for X weeks of Y hours / week at $Z / hour. Certain, the contract can sometimes be ended at any time. However in addition to that, you may considerably depend on a gentle stream of cash over this time so long as the shopper is joyful.
I by no means really understood “threat versus reward” till I began freelancing. Anytime you’re keen to tackle a bit of extra threat, the potential for higher reward will increase. With contracting, your time is straight coupled to your revenue. You possibly can ask for larger and better hourly charges (and it’s best to!), however all potential shoppers will balk at some price.
One of many issues that bothered me about this setup was that the incentives are poorly aligned. If you’re charging hourly and get twice as quick at your job, you’ll earn half as a lot, except you hustle and discover extra shoppers. Sure, you may attempt to cost twice as a lot, however there may be probably an higher restrict to how a lot you may efficiently cost.
You possibly can decouple your time out of your cash by switching to “project-based billing”, which I consider as “company work”. On this state of affairs, a shopper hires you (or your company) to do some challenge for them. This may very well be a small challenge, like an evaluation, or an enormous challenge, like a manufacturing machine studying mannequin deployment. You comply with some worth for this challenge, and away you go. For those who can agree upon a worth proportional to the worth this challenge will convey, quite than the quantity of labor that it takes, you then stand to doubtlessly make some huge cash. Think about you begin to construct out your personal code to make it fast and straightforward to deploy machine studying fashions. You possibly can reuse this code whereas rapidly churning out fashions for various initiatives and cost a excessive quantity for each. The inevitable extrapolation of that is to show your self from a companies firm right into a product firm.
On account of some failed negotiations at larger charges and an rising urge for food for threat, I bought a project-based contract. The shopper was a small startup that needed to generate a time sequence forecast for a bunch of various “entities” every evening. They got here to me with this objective and an outline of the information that they’d. I wrote up a protracted proposal describing how I’d go from auditing their information to deploying a nightly forecast. This proposal consisted of a number of “milestones” that I’d hit (e.g. a working mannequin prototype with attaining X efficiency metric), and a timeline for these milestones. I quoted a weekly price however didn’t point out what number of hours I anticipated to work. On the backend, I calculated my weekly price by choosing an hourly price and estimating my time. This can be a clear threat<>reward tradeoff. If the challenge took me longer to do than I deliberate, then I’d find yourself consuming the price. If it took me much less time, then I used to be capable of obtain the next efficient hourly price.
Consulting
After I consider consulting, I consider paying any person some huge cash to return into your organization and let you know issues. This particular person probably does much less technical work (e.g. writing code) and works fewer hours per week than contractors or agency-based freelancers. However, the marketing consultant might cost a really excessive hourly price to make up for this.
I ended up as this model of a marketing consultant for a pair shoppers. Largely resulting from my experience in sure areas of information science, I had a pair shoppers for whom I would offer “a la carte” consulting on an advert hoc foundation. These shoppers had been free to ask questions through e mail or by organising a gathering. I would offer recommendation, suggestions, and occasional one-off analyses, and I’d invoice the shoppers for my time. Sometimes, I’d signal contracts which allowed to the shoppers to achieve out as much as X hours per week on an as-needed foundation over a interval of Y months.
Consulting like this may be extraordinarily profitable, as you may sometimes cost larger hourly charges. For those who make a reputation for your self and change into an skilled in your discipline (or no less than persuade others that you just’re an skilled), then you may earn some huge cash doing a pair hours of week per shopper for a comparatively few variety of shoppers.
However what does a information science freelancer do?
A lot of the above could also be acquainted to individuals who work in choose industries. Businesses are widespread in promoting. There are extra contractors than authorities employees in DC. However, information science freelancers are uncommon in 2019. Many individuals have requested me what kind of labor I truly did this previous 12 months. My glib reply was, “no matter folks paid me to do”, which was partially true.
One of many causes I needed to do freelancing was that I needed a broader view of the trade. I had actually spent my total (albeit quick) information science profession working at box-based retail firms. Thus, I used to be notably desirous about working with new kinds of companies and new information science drawback domains, notably time sequence modeling. On the identical time, I needed to proceed to do manufacturing machine studying work. My objectives in freelancing guided what I labored on, and so my expertise may very properly be completely different than different information science freelancers.
I used to be fairly profitable in my objectives. I did little or no work with retail firms, however I did discover work with each a B2B SaaS firm and a college analysis lab. The vast majority of the initiatives that I labored on concerned numerous types of time sequence modeling, like time sequence forecasting and classification. I additionally bought to deep dive into some gnarly Bayesian modeling.
Deliverables
My deliverables had been sometimes both jupyter notebook-based experiences or full-fledged deployed fashions within the shopper’s cloud. Deploying fashions was a good way to diversify my technical expertise past what one can sometimes obtain at a person firm. At a earlier job, I had used Luigi for information pipelines, however I bought the chance to make use of Airflow with one in every of my shoppers. Earlier jobs had been primarily AWS (or not within the cloud 😬), however I bought to make use of each GCP and Heroku with shoppers.
Generally shoppers would have an thought for what they need me to construct (“We wish to forecast this factor each evening”). Different instances, shoppers would merely have an issue, and I’d provide you with an answer to this drawback. In both state of affairs, I’d sometimes write up a proposal in regards to the necessities and my method for constructing the machine studying system. This may show a few of my experience to the shopper and provides the shopper an opportunity for enter. Most significantly, this doc could be a solution to maintain us each accountable because the challenge progressed. For instance, requested adjustments may very well be clearly demarcated as “out-of-scope” in the event that they weren’t included within the authentic proposal. On the identical time, milestones that I had promised to hit had been clear.
Uncertainty
Many different information scientists who I’ve talked to discovered the prospect of doing information science consulting terrifying resulting from a specific line of thought. I had the identical ideas initially. The thought course of is as follows:
If any person needs to rent you to construct them an ecommerce retailer, you then no less than know that many individuals have constructed shops earlier than. Ecommerce shops are a identified factor which were constructed many instances. However, if any person asks you to construct them an algorithm to foretell one thing, you don’t know if it’s truly doable to construct an correct algorithm. What’s worse, it can take severe time to get to the purpose the place you may even decide the accuracy of the algorithm!
I believe there are two points with the above line of thought. Firstly, it locations an excessive amount of certainty on “common” engineering work, like constructing an ecommerce retailer. All engineering work brings uncertainty, and we’re nonetheless horrible as an trade at work estimation. Secondly, so long as you talk with the shopper and attempt to suppose issues by in your proposal / assertion of labor / contract / and so on…, then the uncertainty of algorithmic success is simply one other stage of the challenge.
Any Knowledge Scientist is aware of that it doesn’t matter what an organization claims, you gained’t actually know what’s happening till you begin to look by the corporate’s information. Consequently, my initiatives began with an information audit. The information audit consisted of me trying by the corporate’s information, calculating some metrics, making some plots, and different analyses that appeared related for the eventual machine studying mannequin that I used to be requested to construct. The deliverable was usually a jupyter pocket book report. These had been useful for the shoppers as a result of they usually bought a view into their information that they’d not seen earlier than. This course of was useful for me to each get a really feel for the information and area in addition to establish any information points. As an illustration, if an organization at all times overwrites every row of their transactional database, then it will not be doable to construct out historic coaching information!
After the audit, there was sometimes a mannequin Proof of Idea (POC) interval. If doable, the shopper and I’d agree on some efficiency metric forward of time that the mannequin ought to have the ability to obtain. Typically, it’s best to make use of a easy baseline mannequin’s efficiency metric. A flowery time sequence mannequin ought to be capable to beat a easy transferring common. The POC interval is a pleasant alignment of incentives. The freelancer needs to realize good mannequin high quality, in any other case they gained’t get to work on the remainder of the challenge (and subsequently acquire that pay). On the identical time, specific time has been carved out to acknowledge that that is an exploratory interval.
How did I discover work?
Alas, I do not need a hack or silver bullet for this. The vast majority of my shoppers got here by my community. I’m lucky to have labored at a number of jobs that laid folks off and which I give up. This provides me a reasonably large internet of former coworkers. Moreover, the alumni community of Perception Knowledge Science is incredible. I often had shoppers who chanced on this very weblog and reached out for consulting assist. That was tremendous cool.
Whereas I didn’t have a horrible time discovering work, issues probably would have gotten tougher. Identical to D2C firms who face rising buyer acquisition prices as they churn by their greatest clients, I probably would have began to exhaust my shut community and needed to make investments extra closely in each advertising and networking.
What did I like?
There was so much to love about consulting. In my quick time, I bought to significantly broaden my expertise. I labored on new firm domains, new information science strategies, new infrastructure, new frameworks, and so on…
I actually loved setting my very own schedule. I had minimal conferences in comparison with full-time jobs. This allowed me to work from wherever I needed, and I’m now a powerful proponent of distant work. I bought into an important groove working from my bougie fitness center the place there’s a cafe with nice WiFi. After I was working full time previous to consulting, I’d get to the fitness center within the morning, race by my exercise, after which nonetheless need to bathe and commute to the workplace. With consulting, I may exercise within the morning, after which I’d already be on the workplace. Within the summertime, I may work from the fitness center’s ridiculous outside cabanas that look out over the Hudson. As soon as midday hit, the cabanas had been not within the shade, so I’d trip my bike house, clear my head on the trip, make some lunch, and work the remainder of the afternoon, recharged, from house.
Gymnasium Cafe Cabanas “Working”
I met with a lot of individuals whereas consulting. Originally, my mantra was to by no means say no to a gathering. Tech likes to vogue itself as introverts, so many individuals could also be repulsed at fixed conferences. Quite the opposite, I believe it’s an important setup for introverts. Conferences imply that you’ve centered time throughout which it’s a must to socialize. On a regular basis in between will be spent specializing in technical work. I quite preferred this setup. I may meet with folks for espresso, find out about all of the fascinating issues their firm is engaged on, after which go house and “recharge” my social batteries.
On that be aware, it was eye opening to find out about how massive the enterprise world actually is. It’s straightforward to get horse blinders when working at an organization or solely consuming information from sure sources. The true world is at all times a lot bigger than you’d count on, and industries are at all times deeper and extra fascinating than they seem.
Talking of enterprise, I discovered so much about enterprise! I created an LLC, I opened a enterprise checking account, I tracked bills, I despatched invoices, I adopted up on unpaid invoices 🙄, I wrote contracts, I learn too many contracts, and I withheld my very own taxes. My fast suggestions for anyone trying to do all of this merely are:
- Create a fast Single-Member LLC. It’s fairly low cost, offers you a bit of peace of thoughts, and it qualifies you for a enterprise checking account.
- Pay a service to create your LLC (no less than in NYC the place issues are a bit of difficult). I used Northwest Registered Agent. It was $300 complete or one thing. Whereas this may sound like so much, your time is now actually cash, and I’m positive a service can do all of this a lot faster than you may in $300 value of your time.
- Open a enterprise checking account early to be able to put all massive, preliminary bills on the corporate bank card. For instance, I bought like $500 in bank card factors by spending sufficient cash within the first 3 months which was simply happy by shopping for a Macbook on the bank card.
- Use Quickbooks Self-Employed for accounting. It’s low cost, easy, and a very good product. It is going to mechanically match-up photos of receipts with transactions in your bank card. You too can use it for invoicing shoppers.
Lastly, freelancing was thrilling. There’s a thrill to promoting, negotiating, and shutting offers. Additionally, nothing is kind of as motivating as promoting a challenge to any person understanding that it’s going to be as much as you, and also you alone, to determine how the fuck to get this factor to really work. Certain, this was hectic at instances, but it surely additionally made me extremely productive.
What did I not like?
There was so much to not like about consulting. A lot, actually, that I not wish to do it!
Time
After I began consulting, I had this dream of charging some huge cash per hour and dealing minimal hours. I believe I nonetheless have this dream, however my definition of “minimal hours” has modified.
I had hourly contracts for 2 completely different shoppers totaling 30 billable hours / week after I started freelancing. That is too many hours. Most individuals folks would scoff at this and say, “I’m sorry, 30 hours is just too many?!”. In actual fact, we’re simply coming off the newest Twitter battle about what number of hours folks in tech work. I’ll think about the entire arguments moot with out proof of tracked hours and what work was executed throughout these hours.
Let me clarify why 30 hours is just too many. Firstly, keep in mind that that is 30 hours that you’d really feel morally justified in charging $X00 / hour. Typically, these hours encompass so-called Deep Work, and it’s extremely onerous to do lots of deep work day-in and day-out. Even so, there may be lots of different work to be executed as a freelancer. You need to keep your pipeline of shoppers. This entails a lot of espresso conferences with potential new shoppers. I’m fortunate to dwell in downtown Manhattan, so I’m comparatively near most assembly locations. Nonetheless, I sometimes allot half-hour to journey, an hour for the espresso assembly, and half-hour again. Assuming I did billable work proper up till I left for the assembly and instantly bought again to working after I bought house, this single assembly ate up 2 hours leaving me with 6 hours out of an 8 hour workday to do billable work. At 30 hours / week, you’re allowed one assembly per day, assuming you’re excellent at context switching.
Certain, I went on fewer than one assembly per day. However, typically, I had a number of conferences. There’s additionally a lot of time spent writing proposals for brand new shoppers, reviewing contracts, doing accounting, and the entire different issues mandatory with operating a enterprise. Even when I didn’t need to do any of these issues, I’d nonetheless argue that 6 hours / day is probably going the utmost price for good, actual, morally-justifiably-expensive work that anyone can maintain for an prolonged time frame.
It’s true that tech folks routinely go to the workplace for greater than 6 hours / day. Nevertheless, it’s a lot simpler to take a seat by a pair hours of conferences than to actually focus and write code for a pair hours straight. That is why understanding what individuals are doing with their time after they declare lengthy work days is vital. There have been instances in grad college after I labored greater than 40 hours per week, together with time on weekends. Whereas I’ve vowed to by no means do that now that I’m out of academia, this was no less than considerably sustainable (though nonetheless not beneficial!) as a result of lots of my time was spent transferring liquid helium, ready for information to be acquired, or doing different senseless duties like actually watching glue dry.
Given every part above, I’d argue that a super variety of billable weekly hours is one thing like 20-25. With a quantity this low, together with doubtlessly lots of new bills (e.g. well being care), elevated threat (future revenue is just not assured), and no paid time without work, you rapidly understand that you could cost fairly excessive hourly charges to ensure that freelancing to make any monetary sense in comparison with full-time work (and even then, salaries at massive tech firms are astronomical proper now). Many firms will balk at these charges, however it’s best to stand agency. With all of my speak about hourly charges, now you can see why I used to be so desirous about making an attempt to decouple my time from my cash.
Work
When you’ve got not learn Monica Rogati’s Knowledge Science Hierarchy of Wants, then get out of right here and go learn that. I convey it up as a result of my objective was to deal with work in direction of the highest of her pyramid, like writing manufacturing code to repeatedly practice and/or carry out inference on machine studying fashions. There’s considerably much less freelance want for this kind of work than work in direction of the underside of the pyramid, like ETL pipelines and analytics.
For the ML want that did exist, lots of it was misplaced. I efficiently argued myself out of a job a number of instances by dissuading firms from hiring me for ML and as an alternative encouraging them to do some primary analytics or setup an information warehouse.
For the few firms that had a professional ML want, I began to really feel morally unhealthy about constructing manufacturing ML programs for them. In actual fact, I’d stand behind the next assertion for the overwhelming majority of firms:
Don’t rent a freelancer to construct information merchandise in 2020.
As a result of choice bias of firms that had been searching for a freelancer, they tended to be both very early stage startups or had few to zero information scientists. Constructing out a manufacturing ML system requires a lot fucking code proper now. Certain, firms comply with phrases up entrance and that is all enterprise child, however I actually loathed handing over large piles of code + infrastructure to shoppers.
To be clear, I delight myself on the code that I write. I write checks, feedback, and documentation. I attempt to hold issues DRY and modular. Regardless. Many components of an ML pipeline find yourself consisting of customized code. Every part will ultimately break. The information coming into the pipeline will change. The information quantity will develop. Shit will occur, the corporate should keep this method, and it’s going to be a ache to take care of. In spite of everything, ML is the excessive curiosity bank card of technical debt. It’s doable to construct out tooling to mitigate these dangers, however the checklist of required infrastructure is lengthy.
The crux of this situation is that there is no such thing as a Rails for ML. For those who rent any person to construct you an internet site and so they hand you a pile of Rails code, you could find many different individuals who will be capable to take over this code and perceive it. This isn’t the case with ML proper now. It additionally may by no means be the case resulting from the truth that ML code is so tightly coupled with compute / infrastructure.
I additionally bought lots of pushback from potential shoppers who needed this work to be executed in home. On the time, this irritated me (I needed a deal!). In hindsight, this makes lots of sense, and I’ve now come round to the assumption that almost all of this work ought to be executed in home. The programs are simply too difficult proper now.
A counterexample to that is that there are a variety of firms for whom they mainly want easy, one-off information merchandise constructed. For instance, think about a small-scale ecommerce firm that wants a suggestion system. Their scale is sufficiently small {that a} easy suggestion system will suffice, and enhancements to the algorithm won’t have massive absolute impression due the small scale of the corporate.
You possibly can have 2 DS and a 2-year plan for them, however, even when they execute flawlessly, you may find yourself in a spot the place there’s not a lot left to do except (1) there are new DS merchandise to construct or (2) the dimensions of the corporate warrants marginal enhancements on current DS merchandise
— Ethan Rosenthal (@eprosenthal) June 4, 2019
You possibly can think about hiring a freelancer to construct such a one-off suggestion system. That is in all probability the flawed method, although. One college of thought comes from this quote by Josh Wills:
Don’t ever do one mannequin in manufacturing, do hundreds of fashions or zero fashions. For those who’re engaged on an issue, and you could deploy to manufacturing, however you’re by no means truly going to rebuild the mannequin, that could be a sturdy sign that this drawback is just not truly value your time.
Alternatively, one may argue that this method ought to be constructed, however not in home and never by a freelancer. You must as an alternative pay a vendor or another managed service to each implement and keep the information product. The trade has superior sufficient now that a few of these information merchandise are beginning to be commoditized.
This has all led me to postulate that the one customized information merchandise that ought to be constructed and maintained in home ought to both be completely core to an organization’s revenue or be merchandise for large scale firms such that marginal enhancements have massive, absolute impression. Every part else ideally ought to be dealt with by a 3rd occasion which is answerable for upkeep.
Let me now hop down off my soapbox and end this rattling submit. Lengthy story quick, I bought considerably jaded by what information merchandise ought to be constructed at firms, and I began to really feel morally icky about dumping large piles of ML code onto unsuspecting startups. Even when my code did efficiently run day in and day trip, it could nonetheless be onerous for these startups to know that my fashions had been appropriate. Simply because an API serving an ML mannequin returns a 200 standing, this doesn’t imply the mannequin’s prediction is correct. Soiled information may have corrupted the earlier retraining job and ruined the mannequin’s efficiency. I needed a solution to no less than be sure that shoppers had visibility into their mannequin efficiency, which is what I ended up spending 6 months constructing as a startup thought. I’ll save that for the subsequent submit, although.
