For the Ocean Full Container Load market, Flexport offers an “Open Carrier” option where the client books a shipment on something like “FLXT TPEB Ocean Line.” Flexport has had these offerings for many years and the concept is sound, but the details on how we are able to price and fulfill it make a world of difference. Previously, we had teams of operators that would construct these offerings with bespoke processing scripts, sql queries, custom workflows and lots of google sheets. The new internal process we’re announcing today, Automated Offering Creation, replaces that cumbersome manual process with a streamlined pipeline that builds off of the extensive work we’ve put into modeling our contracts and the global network of ocean vessels.

In this post, we’ll dive into the details about what made this simplification possible, how it relates to advances in the Procurement Database and Carrier Space Supply model, and how that makes the Open Carrier vision more attainable than ever.

Procurement Database

The first advance that made this possible is a multi-year effort to build a best in class in-house contract digitization effort. When I started working on the Procurement team in early 2024, we were heavily reliant on third party contract processors. At this point the overall system had serious limitations: only 70% of shipments had an expected cost and of those, the majority ended up inaccurate after comparing it to the final bill. Today, we have expected costs on over 95% of shipments and over 90% are accurate.

One feature that helped us improve our Procurement data is the use of AI to help write custom contract parsers. Ocean contracts come in dozens or even hundreds of complicated formats, often via xlsx files with inconsistent layouts and inherently ambiguous interpretations. It takes a combination of complex python parsing scripts, domain knowledge, and specific clarifications from carriers to get usable rate data. Over the past few years, we’ve built a system that combines custom software engineer written traditional code, expert operator oversight, and AI agent ad hoc code to slowly and surely drive data quality up to a high bar. We’ve iterated through a few distinct designs on this part of the project, and you can read about the platform that powers our latest digitization agent here:

How We Ship Agents Faster with a Configurable Agentic Framework

Carrier Space

The other foundation of this project is the insight that contract data on its own is not enough to really know how much it will cost us to fulfill an open carrier offering. One aspect of the problem is that rate data in the contracts is handled completely separately from two other critical pieces: Allocations and Sailing Schedules.

Allocations

Allocations control how many containers flexport is allowed to book with each carrier. These numbers come in every couple of weeks and are determined at various granularities: sometimes they vary by port or region, sometimes by “service string” (see below for an explanation), sometimes by connecting leg, sometimes by special premium service types, or for specific clients, or a combination of all. From 2021 to 2023 we built an elaborate Allocation Management System for our buy side agreements with carriers, which we now make available to clients of ours who use us for Booking Management.

Shockingly, contracted rate data and allocation data are not maintained by the same folks. Both are complicated constantly-changing agreements between flexport and ocean carriers regarding our bookings on their vessels and yet the lines of communication, job titles, document formats and business logic regarding the underlying data are totally distinct. The natural way this shows up in a tech system is to have two distinct teams build domain specific solutions that stay independent. That was how it worked at flexport for a long time, but we’ve been working for many years to overcome that and this launch is the successful result of that effort.

Sailing Schedules

In addition to knowing how much it costs to move a container and if we’ve been allocated the right to do so, we also need to know what sailings are available. In fact, in order to really understand the set of distinct bookable options we have to fulfill some open carrier offering, this data is the starting point. What routings (ie combinations of trucks, barges, ships, and trains) exist that can get from the origin port to the destination port? The main input to this system is a collection of “service strings” from the ocean carriers. The best analogy for these are bus routes. A typical one may include a dozen or so ports that a ship will stop at over the course of a couple months. Maybe a few ports along the coast of Asia, then cross the Pacific to the west coast of the US and Canada. Most carriers send one ship per week for their major strings and often a collection of carriers coordinate via “Alliances” so that multiple carriers can offer space on the same vessel. Flexport maintains the world’s most advanced collection of sailing schedule data and makes it browsable for free at atlas.flexport.com .

Here you can see an example vessel, the OOCL ASIA, traveling on CMA’s service string BOHAI from Tianjin to Qingdao to Shanghai, then California and British Columbia.

We combine the concepts of Atlas Sailing Schedules, Allocations, and Contracted Rates from the Procurement Database into single bookable units of space that we call a Carrier Space.

We first used this data model with early versions of the Procurement Database to power our Fulfillment Management Tool and Boat Assignment Engine. This was deployed internally in 2023 and has become a core pillar of how we fulfill ocean shipments. In the below screenshot from a March 2025 announcement you can see the combination of Allocation data (25/100 teus assigned), Sailing Schedule data (18 days to go Ningbo>Norfolk>Columbus) and a failed attempt to lookup rate/cost data (Calc cost: UNKNOWN). These days we have far fewer UNKNOWN costs because our shipment rate coverage is over 95%.

Bringing it all together

For the first time we now use that same integrated world model to answer the question at the core of this problem. If we were to fulfill some Open Carrier Offering, how much would it cost? First we determine which Carrier Spaces can be used, then we get the cost for each one, then we look at past data to build a blended composite cost that powers our Open Carrier prediction.

The overall flow is much more streamlined

Below, you can see for an Open Carrier offering from Shanghai through LA to Chicago, we have three different carriers, 4 service strings, 6 different departure dates, 4 different Allocations, 6 different booking references, and 4 different costs. And notice neither departure date, nor service string, nor allocation, nor via port is enough to determine the cost; cost can depend on all of them and that data is spread across all three systems.

There’s still a lot more to be done, but by building on top of sound integrated fundamentals and ensuring the entire flexport system uses a shared system of record for critical information like where and when ships are sailing, and how much it costs us to move freight on those ships, we are making steady progress towards our vision of a brighter world via seamless global commerce.

From Contracts to Catalog: Automating Ocean Freight Offerings was originally published in Flexport Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.

The internet lets you have the best instructors, but what we really want (and can do) is get rid of instructors

Originally written 2016, first published May 2020

I think there are essentially three ways education can be improved:

• Motivation
• Quality of consumed lessons
• Moving beyond consumption based learning

Motivation

A kid who wants to learn is far easier to teach than one who doesn’t. Arguably, the most challenging and important work teachers and administrators do is pushing students to be motivated and buy in to the learning process. This involves classroom management, individualized education plans, parental outreach, and plain hard work.

Technology can surely help on this front by providing tools for communication and planning. Companies like Remind and Edmodo and others are raising huge sums of money[1][2] with the promise of helping teachers with this important challenge.

But while adding messaging functionality and modern interfaces to educators’ toolkit can help, I don’t think technology is the silver bullet here. We need better student / teacher ratios, more community outreach, awareness and training for cultural differences and learning disabilities. None of those will be quick and easy wins, or cheap, but they are all crucial if we as a society wants an educated public.

Quality of consumed lessons

The traditional methods of formal education include a variety of ways to transfer knowledge from an expert to a novice. In the 20th century, this was primarily textbooks and in person lectures. Recently, online videos have surged in popularity as a medium through which courses are taught.

I think technology enables enormous room for improvement. Because distribution of bits is so incredibly cheap and easy (compared to speaker tours and even physical books, not compared to 0). Everyone can theoretically access everything. But how to bubble up the best is still a hard an important question, that deserves a thoughtful strategic approach.

A related role of technology is that everyone can be a contributor. Put Coursera, Udemy, and Youtube on a spectrum to see the power of crowdsourceing.

But, in the mean time, let’s use a human curation approach and single out outstanding content to be consumed. One great instance was Khan Academy videos. Sal Khan is probably in the top 1% in his skill at explaining concepts. My guess is that the average teacher is in the top 10%, but that means that for an enormous number of kids, consuming an explanation from Sal Khan is better than consuming an explanation from their teacher.

Bill Gates recently wrote a post entitled “The Best Teacher I Never Had” about Richard Feynman. Here’s a video narrated by the nobel prize winner:

Imagine what it would be like to have Richard Feynman as your high school physics teacher. That is achievable. There’s no reason why everyone shouldn’t consume the best version of any lesson.

To a certain degree, textbooks already fit this model: there are a handful of calculus textbooks that collectively are in 90% of classrooms[1]. But since an engaging lecturer is so much better than a textbook, we should strive for everyone who ever sees or hears a lecture to get the best version of it.

Moving beyond consumption based learning

Paulo Friere describes what I call consumption based learning as follows:

Instead of communicating, the teacher issues communiques and makes deposits which the students patiently receive, memorize, and repeat. This is the “banking” concept of education, in which the scope of action allowed to students extends only as far as receiving, filing, and storing the deposits.

That is from page 58 of his book, “The Pedagogy of the Oppressed”

The Future of Learning

The primary difference between educational lessons of the future and those of today will be interaction. The basic unit of learning 20 years ago was a chapter in a textbook. Now, on the rising new ed tech platforms (Khan Academy, Coursera, Udacity, Udemy, etc) the unit of learning has become a video. These platforms have a couple of hacked together quizzes that sit on top of the videos, but that’s the very very beginning of what interactive learning could be.

a few examples include

geometer’s sketchpad, phet physics simulations, khan tangent line, desmos (tool that could be used for tons of different lessons), open gl formulas — some of which is super high quality consumption other is genuine interactive

In that vein, I’ve created a proof of concept for teaching k12 math through interactive lessons, focusing on exponent rules.

Here’s the github page, which includes a demo video. Sometimes it’s available on the App Store as HandsOnMath, sometimes it’s not.

https://github.com/mlipman/HandsOnMath

I think a really important test for these types of lessons is: If I were to ask you to export it to youtube, how much of the instructional value of the lesson would be lost. Right now, that number for the best courses there are is shockingly low. The only real exceptions are computer science courses which because of the interactive nature of programming have been embracing interactive lessons since early on. Other than that, almost all courses are basically videos.

With the rise of smart phones we are entering the first time in history where the audience for interactive lessons is reaching the billions. if you believe that and audience and demand for a product must exist before that product gets made, then it makes sense that until a few years ago, almost no one was producing interactive lessons. but we’ve now reached a tipping point where the supply of smartphones has enabled the future of learning.

What does it look like

A menu of lessons. most of which are small bites, single key ideas. on their own, if you were to hear someone else explain the idea to you (which you shouldn’t) it would take between 1 and 5 minutes.

But you’re not having it explained to you, you’re discovering it yourself via an interactive lesson.

No Spoilers

One idea that came up in a course I took on the Pedagogy of Math is that having someone explain a cool or insightful academic idea to you can actually be detrimental. It deprives you of figuring it out yourself, which is both fun and a boost for your self-confidence, but also helps you remember it.

One of my favorite school memories is from 5th grade, was when me and three others, with the right nudges from our teachers “discovered” prime factorization. I think the insight was that all numbers are uniquely defined by a set of little numbers you can multiply together to equal the big one, and all of the little numbers are the primes we’ve been learning about.

I remember looking at the notebook paper tuning high double digit numbers into just 2s and 3s and 5s and thinking wow, I can’t believe we figured this out all on our own! If me and my buddies were around in Ancient Greece, we’d be famous by now.

Of course we didn’t figure it out on our own, but a system was built that let us get there and that made all the difference. There’s an adage: “Tell me and I forget, teach me and I remember, involve me and I learn”. I truly think that when possible, you should try to enable self-discovery and avoid teaching ideas that students can derive on their own.

What’s on the menu?

A list of topics that seem good for this type of self discovery

Exponent Rules

Stoichiometry

Position vs Velocity vs Acceleration

Other sources

Common Core

Big list of metal models: See bottom of that article for about 50 concepts, each could be its own interactive lesson.

A Product Vision

If you wanted to turn this into a consumable product, here’s one vision of how the platform could work:

The platform itself, Hands On Learning, is a catalogue of lessons. some can be developed manually in house or commissioned. but real power comes from crowdsourcing. a few key ideas

1. anyone can create a lesson. simple as creating a view controller or a single web page. can be pull requested into the bigger project, with a receiving address for money. perfect for the rise in cs classes, where students (who are often exceptional at at least one other class) need to design and implement a final project. why not make a lesson?
2. compensation. lessons and help would be paid for. maybe if low income you can get an initial grant. others pay for credits to start
3. you can tutor for credits, credits are essentially money, so you can buy lessons and tutors with them. you can earn credits by helping out (async or sync) or building lessons, or you can buy them. maybe buy to get started, then try to earn yourself.
4. paying out: If the platform sends you there, 50/50. If creator gets you there: 100% to creator. (Based off of Udemy’s model).
5. Curation: perfect place for data augmented curation. AI and machine learning have a huge role to play in determining which lessons are best. But comparative advantage over other curriculum analytics tools is not algorithms, it’s the data. you can capture much more and much richer data that using it to test human created hypotheses about learning will be an enormous advantage.

A continuum of interactivity

When I wrote this in 2016, I saw interactive lessons as a different type of learning vessel from video or spoken lectures. But certainly all instruction lives on a spectrum of how actively the learner is participating. For example, there are two lessons linked below [1][2] that are scrollable visualizations. In some ways, they can be almost fully encoded by an article with images, but the experience of reading it is so different because as you progress through the lesson, each image has many different phases, and you can sit on the part that is most crucial to you. It’s actually a strikingly radical difference from scrolling through a static article with images.

—

Miscellaneous Notes

Four key ideas from stanford professor candace thille on the opportunities from online education

1. world wide distribution and 24/7 access
2. interactive lessons
3. global classroom. pick best ta in the world. stack overflow > office hours
4. collect and use data to improve (address privacy concerns, but don’t let anyone throw out the gold. stakes are too high for students)

—

Students hate textbooks, they see them as boring, utterly useless to their pursuit of passing their classes, and exorbitantly overpriced.

—

high school lecture classes (to me) weren’t as bad as college lecture classes, mostly because I had a personal relationship with the teachers and that contributed to more motivation and more buy in.

—

Random comment I made on an article about using slack to teach a college course:

I think the best solution would feel as lightweight as slack to students, have searchability and tagging like stack overflow, and allow for efficient processing by a team of assistants like how customer support teams process tickets.

—

Contrarian view point: “teachers” should be replaced with coaches and counselors who are not subject domain experts and not necessarily paid above median wages.

Designing for someone with a little motivation is still good and will translate to those with low motivation. if learning doesn’t suck, more people will be into it.

—

Other cool data visualizations (2020)

https://wanderwhim.com/html/nba-season-history.html

http://www.r2d3.us/visual-intro-to-machine-learning-part-1/

For some notifications, if you don’t act on them when they initially show up in the lock screen, you don’t get a second chance. They probably aren’t completely vanished, but by opening your phone, you are acting on them, which you may not want to do right then. In many cases, I do act on them then, which derails my quick task into an extended session.

That’s one reason why I think voice interfaces (siri, alexa, hey google, etc) have been catching on. I still think they in general offer way less rich user experiences, but they allow for people to complete tasks without throwing their inbox in their face.

tldr: Collective bidding with enforced price drops for software is 💯💯💯

Imagine a enterprise software company is considering building out a new product. It will cost them $100,000 to build. They think the best price is about $30 per month per customer. Let’s say all customers are likely to use it for about 3 years, so the lifetime price paid per customer is $1000. The company thinks they will get more than 100 customers at that price, so they move forward and build it.

That’s the base case: $100,000 cost, $1,000 per customer, estimated number of customers: slightly more than 100.

Under adaptive pricing, the price per customer would go down by a factor of 2 if 4 times the number of customers committed to the lower price.

If 400 customers commit to pay at least $500, then the price drops for everyone to $500 and total revenue goes from $100k to $200k.

If 1600 customers commit to pay at least $250 (less than $10 per month instead of $30) then the price drops to that for everyone. Revenue goes up to $400k.

If it turns out that stingy college students are actually a great demographic, with 10 million who want the product, but they’re only willing to pay $0.10 a month, revenue would go up to $3.3M!

Technical Details

The price for a piece of software should be determined by a 30 day auction, where all potential users can enter a bid of how much they’d be willing to pay for full use of the product. If the cumulative quantity of bids at a lower price is high enough such that total revenue is significantly higher than the original scenario, then the price for everyone drops to that lower price.

All bids are of the form: “I would pay x amount or less for this product”. That means as you consider lower prices, the cumulative amount of customers increases. For a given price (p) and the corresponding number of customers at that price or lower (q) you make the following comparison: (Using base condition of 100 customers at $1,000). If q/100 ≥ ($1,000/p)^2 then the price drops to p. You do that over and over for smaller and smaller prices, until the condition no longer causes p to go down.

Assumptions

There are two main assumptions about the supply and demand of the product that if met, would make this pricing strategy work.

Extremely low marginal cost (not average cost)

This is true for most software. Let’s consider a single version game without an online mode that’s distributed on the app store. Once it’s shipped, it doesn’t cost the developer anything for the next person to download it. That is an extreme example, since most software requires more cpu time and data storage as the number of users goes up, but the true marginal costs remain very very small.

Customer support is a different good all together and would have to be priced separately. Also, this pricing model doesn’t apply to many of the software enabled companies that sell physical goods or services: Amazon, Lyft, AirBnB, etc. Adaptive pricing would work best for products like YNAB 4 and Sketch3. But I think most if not all software development would benefit if this type of pricing model became common place.

Here’s a litmus test to differentiate between whether the software itself is the asset being sold, or something else: If Congress passed a law tomorrow requiring all code to be entered into the public domain, what would happen to the stock price of the company? Sketch3 would no longer have a way to make money, so their stock price would go down. That’s how you know they sell software. Companies like facebook and amazon would be fine. In fact, they spend so much hiring software developers, that a purely open source world would probably help their bottom line. They don’t sell software, they use software to sell ads and goods.

No steep cliffs in the demand curve

Companies that sell goods are out to maximize profits, not maximize reach or impact, and the pricing model must reflect that. It does. As long as the demand curve is smooth, adhering to this model will guarantee that total revenue either remains the same or increases.

Strictly Better

An optimal pricing model would be “strictly better” than the alternative, ie a constant price of $1,000 per customer. This model gets pretty close. There are two instances that deserve further attention.

A Steep Cliff

If 390 customers would be willing to pay $1000 or less and only 10 would pay $500 or less, then adopting adaptive pricing would lose you a lot of money. Your revenue would be $200k when it could have been $390k. So if you have good reason to believe the demand curve drops off at a specific price, then price your product just before the drop-off. But for most products there’s no reason why the demand curve would be like that.

Breakouts

Another concern that people might have is something like, “If 100x people want to buy it for $1,000 dropping the price is throwing away money!”. The thinking being that if demand is much higher than expected, price shouldn’t go down, it should remain high, or go up!

But this is only true if the supply curve is high and increases with quantity supplied. Demand doesn’t “go up” it shifts right. And if demand shifts right, quantity supplied should go up. So while I appreciate the sentiment, the model adequately rewards breakout products. If the demand is a lot higher than expected, total revenue will be a lot higher than expected.

Price Discrimination

Perhaps the most important factor in making money selling software today is the ability to price discriminate. This model purposefully does not allow for that. Software as a service companies spend a huge amount of effort and money on getting their customers to pay close to the value that they’re receiving. That’s great for the companies but bad for consumers. And in my opinion, the effort and money spent on that effort represents dead weight loss. There are certainly some arguments for ways in which customer acquisition can improve products, but the modern startup compares CAC to LTV in the same way classical micro economists compare marginal cost to marginal benefit. If you step back and think about that, it’s utterly ridiculous.

Software, once written, has a distribution cost, ie marginal cost, of $0. It’s time for it to be priced that way.