Ask HN: Did your startup come out of academic research?

Where did it come from, what was it, and how did it turn out?

I'm not sure if this counts, since it was sort of an aborted startup. But in any case:

As a spinoff from some machine-learning research on estimating different properties of graphs given incomplete information on other properties, was working on a service that would estimate travel times at different times of day given only fairly high-granularity crowdsourced information. You'd put in a fairly course route for your commute (e.g. I-85 at [exit] to I-285 W, exit [blah], take road west to [major intersection]), then enter data like "left at 7:35am, took my saved route Morning Commute, and it took 22 minutes". From many such users, hopefully with overlapping commutes ;-), it'd back out estimates of how long different segments and intersections between segments would take at different starting times, along with some nice interval estimates ("90% of the time, this route around 9am takes between 9 and 23 minutes, median 17"). Possibly other error-type estimates also: a big interest was that a curve showing mean travel time is missing a lot of the story, since backups due to accidents are very irregular (a route might take you 10 minutes almost every day, but 90 minutes sometimes... how frequent are those delays, and how long?).

It failed, due to a mixture of not enough time spent ramping it up, competitors, and a misprediction on technology that would've sunk it anyway. The not-enough-time was that we were building it as a side-project while in grad school. It wasn't even really anyone's thesis, and had no full-time employees, so it really was a side project, which made things slow. Fortunately or unfortunately, the bigger problems were that competitors and misestimates of technology would likely have sunk it anyway.

Google Traffic launched pretty early into development. It didn't do nearly what we were planning: even today you can't use it to get nice traffic-time curves at different times of day and different days of the week. But it did bite out a chunk of low-hanging fruit, covering both the common "what's traffic like now?" use case, using real-time sensors we don't have access to, as well as a generic take on the time-estimate problem (the Google Maps "up to N minutes in traffic" thing).

The technological misestimate was that we thought there was a bigger window before ubiquitous GPS with easy upload of data made this kind of inference from courser data unnecessary. We were working on this around 2006, and expected it to be maybe 2010 before that kind of thing was at all common, but it began becoming reasonably common in 2007, and exploded with the iPhone 3g and first Android devices in 2008. Obviously if you can take direct readings all along a route, you don't need any fancy estimation to back out segment estimates from whole-route timings.

Unfortunately, I don't know of any website where the kind of data we had wanted to make available is actually available currently. Can I go somewhere, put in "I-280 from CA-85 to I-380", and get a 24/7 curve showing me how estimated travel time varies by day of the week and hour of the day, ideally with some interval bands?

edit: Partly answering my last question, it seems the city of Atlanta actually does give you some traffic curves. They don't show a whole lot of data, but they do provide some aggregated curves for weekday rushhours as of about a year ago for various segments of the freeway system: http://www.georgianavigator.com/histdata/trip.shtml

Two did, yes. One was ArsDigita, which basically came out of Philip Greenspun's Ph.D. work on database-backed web sites for online communities. The second was a compiler company that spun out of a research project in program analysis and compiler design. In the second case, the product evolved throughout development, so it was inspired by but ended up looking very different than the original research.

Before entering university, I had an applied research project that generated visitor passes by performing OCR on identity cards. The idea was to use them for protected locations as a better alternative to leaving your precious identity card at the security post. It worked but never went anywhere because I didn't know how to do very much algorithmic analysis and design work and didn't have the confidence to sell what I had simply put together by experimentation.

Now I do a lot of algorithmic analysis and design but see my research work as too difficult to commercialize. Ironically, my current startup idea doesn't have academic research behind it as a competitive advantage... It's just a service that I know people want.