The Berkeley Science Fellows Program: On campus professional development opportunity for postdocs



The mission of the program is two-pronged‐ provide career transition opportunities to postdocs by helping them experience working with startups; 2. Helping startups solve scientific problems and reach their milestones while being incubated on campus.

Our team

Naresh SunkaraEvangelia VamvakaEduardo González GrandíoAriani Wartenberg, and Gaia Andreoletti


The core function of the Berkeley Science Fellows Program is to connect postdoctoral researchers with startups in the UC Berkeley entrepreneurial community. This will provide valuable real world working experience to postdocs while helping the startups benefit from the scientific expertise of the postdocs.

Berkeley Science Fellows will offer their advice to startups to solve specific scientific challenges including product development, process trouble shooting, grant writing/reviews, experimental data analysis, and presentations, to participate in discussions in brainstorming sessions to develop start up ideas.

The Berkeley Science Fellows Program will be responsible for recruiting postdocs (after an interview) and present them to startups at SkyDeck. Science Fellows will be given an opportunity to select the startup they would like to work with as well.

Qualification: Candidates must be a postdoc at UC Berkeley to be eligible to participate.

Time Commitment

It is understood that postdocs have a primary commitment to their research at UC Berkeley.  Science Fellows will need to commit 4-5 hours a week to assist the startup. The hours are flexible and may include evenings and weekends. Please note that the Science Fellows are not compensated for their volunteer assistance.

Because postdocs will gain valuable experiences while working in a fast paced startup environment, the selected postdocs are expected to deliver on their agreed-upon time commitments. To this regard, the Science Fellows will be intermittently contacted by the Program leads to assure that all commitments are being met.  Additionally, there will be monthly participant check-ins to solicit feedback to improve the Berkeley Science Fellows Program. 

Benefits to postdocs

- Gain experience working with startups.

- Get exposure to potential employment opportunities

- Add work experience to your resume

- Receive a letter of recommendation from a startup

- Serve on the Science Fellows Advisory Board (after completion of internship)

- Receive satisfaction of helping a startup succeed

Benefits to startups

- Access to world class talent (free of cost)

- Future Employee pipeline

- Help with troubleshooting specific problems

- Help with grants (review/write)

This is an unpaid (volunteer) program.

Please use this link to register as a Science Fellow:

If you are a startups trying to recruit postdocs as Science Fellows, use the link below to register: 

 Skydeck Startups recruiting Berkeley Science Fellows

Current postdoc consulting opportunities:

Skydeck Startups

SuperAnnotate AI

We provide super-fast and pixel accurate image annotation services. 

 Problems we need help with

1. The problem we would like to address is simple but could have quite a huge impact in the entire computer vision society. Somewhat similar study has been publish in the recent CVPR in Salt Lake City. The article can be found here:

Our goal is slightly different that the one presented in that paper. We would like to address the problem of bounding box annotation vs pixel accurate annotation. for object detection tasks. Obviously, pixel accurate annotation will provide better object detection. However, one can say that pixel accurate annotation is more time consuming. On the other hand, with SuperAnnotate tools, the time we spent on pixel accurate annotation and bounding box annotation is comparable. Hence, very interesting research problem arises: how many times less pixel-accurate annotated data do we need in order to achieve the same object detection accuracy (IOU)? 

The outcome of this experimental research article could be that the industry should not spend time on bounding box annotation, which is the most common annotation technique used in Automoous Driving and other industries. 

2. There is one more research problem related to a real-time pixel-accurate semi-supervised video segmentation/annotation algorithm that we have been developing at the moment. Please contact me for more details.

You will have 2 ML engineers from SuperAnnotated who will be involved in these research projects. By the end these research proejcts we expect to write ICCV or CVPR article, or just publish it on arxiv. SuperAnnotate will be a Gold Sponsor for CVPR this year, so as a sign of thankfullness we will cover the trip and conference expanses for the post-doc researchers. 

Skills needed

General skills in Machine Learning and particularly in Computer Vision. Good programming in Python. More specifically, we will need motivated researcher who have done some research in Semantic Segmentation, Object Detection tasks. Good understating in convolutional neural networks. 


70% of the projects fail according to Gartner research, mostly because of the project managers and project assistants performance, so basically it's because there are a lot of manual unautomated tasks, and they have not enough time to track and automated everything, besides, the online collaboration is bringing another issue, which is the decentralization of the information. On the decentralization issue, people need to have technical skills in order to make things works integrally. On the PM roles performance, their job basically is to manage every step in the project execution which is divided into three main phases: 1. Planning - 2. Tracking and Suggesting - 3. Reporting and learning. The problem is today these roles are managing multiples projects at the same time, they are using multiples tools which decentralize the information, and at the end, they need to collect all this data and act on real time, which is pretty much impossible. The project manager role needs to be focused on what it matters on the real-life of the project execution and an AI needs to take the project execution process(1. Planning - 2. Tracking and Suggesting - 3. Reporting and learning.) on the digital side. Basically, the AI need to remove the project assistant role and bring a powerful centralized way to the project managers for managing projects more efficiently. By reducing the project fail execution to 20%. (50% of margin improved).

Problem that our startup is solving

1. AI-machine learning which can predict project manual tasks - then we need to create and train models that can process over 55 Million of data points every month. 
2. Data-Science expertise to create algorithms which can execute these operations faster without consume so many cpu resources. 

Skills needed

Machine Learning (sklearn, tensorflow, pytorch, etc.)
Business Development 
Business Relationships 

Obviously AI

Software companies want to predict customer’s behavior from their internal data. Predicting customer behavior requires Machine Learning (ML) which is not feasible for all software companies as many don't have a data science team or cannot scale one. Therefore, we are building a web platform that provides simple drag and drop interface to predict customer behavior. Enabling non-technical folks to run machine learning and get data insights instantly. No programming required.

Problem our startup is solving

1. What is the best way to visualize predictive data and deep dive into the same.

2. Building a parallelized preprocessing pipeline from SQL databases to ML models.


 Skills needed

Data science, Python, JS, Apache beam, Apache spark, Dev-ops, Systems

Squishy Robotics

Problem that our startup is solving

Squishy Robotics develops stationary and mobile sensor platforms for disaster response and remote monitoring. Our robot platforms have the unique ability to be deployed by drones or aircraft, thereby reducing the exposure of human operators for improved situational awareness.

Skills needed

Chemical Dispersion modeling, Analytics/Computation, Vision systems


Pow Genetic Solutions

Problem that our startup is solving

While biomanufacturing may be one of the most technologically complex and profitable industries, the underlying production process, batch fermentation, has not changed in decades. Owing to the high risk of contamination, continuous flow bioreactors are considered unreliable and are rarely used. Replacing the traditional batch system with a continuous system would both significantly improve process economics and enhance production yield. However, continuous bioreactors are notoriously sensitive to microbial contamination due to their long operation time. We provide a cutting edge, genetically engineered solution for fermentation scientists, dramatically increasing yield by treating and preventing microbial contamination.

Skills needed

 Molecular Biology, Microbiology, Synthetic Biology, Fermentation


Problem that our startup is solving

Wavelength is an advanced AI solution company to improve safe driving of today and autonomous driving of tomorrow. We create efficient deep learning software to enable ADAS and driver monitoring functions on a single device and provide big data analytics for OEMs, fleet, and insurance companies.

Here is the company pitch at Plug and Play:


Problem that our startup is solving

DeepScribe is an artificially intelligent medical scribe. We use AI and NLP to tackle the multiple hours of documentation physicians have to do daily by simply listening to the natural conversation they have with their patient. We are solving the burden of clinical documentation in healthcare today which is leading to dangerous levels of physician burnout, inefficient hospitals, and unsatisfied patients.

Skills needed

Experience working with/developing on open source speech models such as Kaldi (beamforming knowledge is a plus)

Experience in the latest in natural language processing methods/developments (word vectors, LSTMs, Seq2Seq, Information Extraction, Named Entity Recognition)


Problem that our startup is solving

covexo makes it easier and faster to build cloud-native software with containers and serverless technologies. The core product of covexo is DevSpace CLI, an open-source tool that lets developers connect their local workspace with any Kubernetes cluster to build, test and run software in an environment as close to production as possible.

covexo's commercial product, DevSpace Cloud, additionally helps software teams to build software collaboratively, to iterate faster and to easily deploy cloud software with confidence.

Skills needed

Postdocs with research interests in the following or related fields would be a good fit: 

1. Information systems research (developer team workflows, software development methodologies, distributed development teams, prototype-based communication, DevOps)
2. Computer science research (DevOps, software testing and CI/CD, cloud computing, container and serverless technologies, software-defined infrastructure)
3. UX & Design

4. Business-to-Developer Marketing & Enterprise Sales 


Problem that our startup is solving

Founded by gaming industry experts, Korro is focused on helping game developers monetize their games using the blockchain. Game distribution, as well as the monetization of in-game assets (including a secondary market for player-to-player trades), can be effectively handled via the blockchain and an associated development kit, avoiding the fraud issues that plague this space


Problem that our startup is solving

Telos.AI is a next-generation visual collaboration tool for the workplace. It’s like an infinitely-sized digital whiteboard in the cloud. Minimal drawing skills are required because Telos.AI can transform any word into visual pictograms (glyphs) on the canvas using our AI drawing assistant. Additionally, users can drag and drop links from other tools (like Dropbox, Tableau, Slack, Jira etc.) and these appear as real-time widgets on the canvas - i.e. you can glance the status of the Jira ticket without having to visit Jira. Telos.AI drawings are more like visual knowledge maps than static drawings. Our primary use case is for knowledge workers to explain any project, conceptual or concrete, more expressively, efficiently and completely than any other method of documentation possible. Expressive and accurate explanation of projects is key to enterprise productivity.

Skills needed

To make CogniDraw useful and better than existing drawing tools, we wanted to make it easy to express ideas at “the speed of thought”. Right now, CogniDraw lets you type in any word into our 'global search bar', and using our novel synonym-based NLP behind the scenes, returns a range of glyphs that represent that word. These glyphs are intended to have a primary meaning, and visually they have a sketch-like, hand-drawn style that you might find on a whiteboard. 
CogniDraw even lets you represent abstract concepts and multiple word senses: e.g. searching for 'lean' returns glyphs corresponding to the literal sense (a thin person, in-shape, etc.) but also glyphs corresponding to the more business-specific sense of the word 'lean' as it relates to the 'lean methodology' and similar agile development-related concepts.

We have built an extensive library of well-tagged icons, or what we call glyphs, which we consider a major asset. We use humans to curate the icons and we intend for the time-being to use “human in the loop” AI processes (i.e. curators) until such time as we can fully automate, by which we mean even use AI to draw the icons (which we aim to do eventually). 

There are a few projects that might be of interest:

1) Glyph recommender - “Glyph2Vec”
We're building a glyph recommender that can suggest glyphs to be used on the canvas simultaneously as the user is drawing something. We want the recommender to do something beyond just recommend glyphs that are synonyms of the glyph that was just placed on the canvas. We envision a recommender that works in both the spatial and time domains:
- Spatial: Suggesting glyphs based on where pre-existing glyphs are co-located on the canvas. E.g. figuring out how to 'cluster' glyphs together based on both their semantic content and their location, and then recommending glyphs that relate to that cluster in some way.
- Time: Suggesting glyphs based on the order in which the previous N glyphs were placed. 

This is where the parallels with n-grams come in, where you could imagine taking into consideration, say, the last 2-3 glyphs placed on that specific canvas (some experimentation needed). Just imagine if the canvas could 'predict' what glyphs you wanted to use next -- that would make for a really fast, expressive, seamless drawing experience. 

There are a few ways we can think of for approaching this problem — one solution might be purely textual (e.g. via RNN, LSTM or similar) utilizing the primary description and synonyms associated with glyphs. Another approach might be to go the CNN route and utilize both the visual content of the glyphs as well as the descriptions. But there are probably a bunch of methods we hadn’t even considered.

2) Enhancing the business/development lexicon
Our goal is to expand the vocabulary of the Drawing Assistant to be as comprehensive and accurate as possible. So we also scrape the web for sources of contemporary business words/phrases/concepts and use other sources like the OED. We then created ~2000 hand-drawn glyphs for these phrases and tagged with ~15K words. The expansion via AI results in ~30K words. Over time we want to make the expanded vocab much larger, but we are keen to avoid noisy or incoherent words. So we’re looking for someone with NLP experience to help us navigate this task. Ideally, we would incorporate topic modeling via NLP.

3) Style transfer to change glyph styles
We want to expand our glyph library to include different glyph ‘sets’ that are each consistent in terms of style. We’re interested in seeing if style transfer via GANs might be possible on our current set of glyphs to change the style, a bit like here: E.g. one set might have a very ‘technical’, minimal straight-lines aesthetic (e.g. for sales or business diagrams), and another glyph set might be more cartoonish (for teams in brainstorming sessions) etc. 

4) Developing and researching the user experience (UX) of Telos
Aside from the CogniDraw/NLP projects, we’re looking for someone to help us research and advance the user experience of our tool. Ideally, this person has experience in researching creating tools for knowledge management, or is interested in the idea of what “white-boarding in the cloud” might look like. We want to make sure we’re designing a UX/UI that is easy to use, is simple (aesthetically) and truly does facilitate “drawing at the speed of thought”. Ideally, this person would be able to play around with our MVP for a few minutes and immediately be able to tell us what we’re doing wrong :-) Or have interesting ideas as to what we could do next, or what untested assumptions we’re making in the design. Note that our platform is UX agnostic and we have thus-far built a 2D web-based drawing UX. Our goal is to build an AR interface using technology like Magic Leap (whom we are talking to).

Abalone Bio

Antibody therapeutics discovery for GPCRs

Problem that our startup is solving

1) Yeast engineering and yeast-based functional selections and assays 2) Bioinformatics analysis of NGS data regarding antibody hit enrichments 3) Antibody library design 4) Bioinformatics of NGS analyses of antibody library construction 5) Mammalian stable cell line construction 6) Molecular structure prediction/modeling of antibody and GPCR structures.

Skills needed

Molecular biology, yeast engineering, mammalian cell engineering, yeast cell based assays, mammalian cell based assays, antibody library construction, bioinformatics, flow cytometry


Spext is Photoshop For Voice. We make editing and production of voice media (podcasts, audiobooks and radio ads) easy.

What is the problem we are solving?

With rise of Alexa devices (100M+), Airpods (40M+) and CarPlay/ AndroidAuto,  the demand for voice content is growing rapidly.

However, production of voice content is expensive ($150-$4000/ hr) & time consuming because:

1. You have to learn waveform based music production softwares such as GarageBand/Audacity.

2. In case the recording is not right, you have to call artists/ guests again to re-record segments.

How are you solving the problem?

Very few people know how to use GarageBand, but everybody knows how to use a text editor, so Spext has built a "Text Editor for Voice." Here's how it works:

1. Upload a recording and it will automatically be transcribed to text.

2. Spoken words and the transcript will be synced very accurately.

3. This means you can select a sentence of the transcript, hit delete and it will cut the corresponding media.

4. You can highlight a sentence and add music in the background. Ctrl-C Ctrl V, Ctrl-X Ctrl-V also works to rearrange different conversations into a cohesive story.

5. And finally, here's something awesome: You can type in new words into the transcript and it will synthesize them in the speaker's voice. (No need to call the guest back to the studio again!)

What kind of Science Fellows are we looking for?

Fellows with backgrounds in ML, Speech Recognition, Synthesis, Deep Learning, GANs, Signal Processing.


1. Creating language independent custom TTS

Spext is working on custom text to speech (TTS) which can be trained by 10min of data in one language ( for e.g. English).  This TTS has also the capability to generate speech audio in other languages. The big use case is voice overs and dubbing. 

2. Smart audio insertions and replacements in a recording

This project is focused on the use case of editing a voice media by inserting one or two word, replacing a word with another spoken word ( without recording ) such that the insertion feels natural.

This will help in correcting the pre-recorded videos/audios without re-recording and also can generate multiple versions of any recordings.

3. Synthetic Media

We define synthetic media as “Algorithmically generated audio that modifies or enhances a recording.”  Unlike TTS, the goal is to maintain audio features of the speaker/s in the recording and change 1-2 words or small sentences in them.

Spext is also working on technologies which are suitable for generating synthetic voice media such as jingles, songs, etc. Which can also capture the tone and style of a speaker and later generate the audio in a specific speaking style.

4. Voice Conversion

If your research interests are around Text to Speech and Voice conversion, we would love to explore opportunities with you.

5. Synthetic Video

We are happy exploring how synthetic audio and video can be combined to create complete experience, which is algorithmically generated. Which will have capability of scaling programmatically and dynamically.

6. Identification of fake audio and Video

With all the synthetic media around, Spext is also working on the identification of synthetic media from real one, this will help in securing digital media rights and will create safe mechanism to work with this technology constructively


Problem that your startup is solving

PredictEV is building a blockchain powered social network, that will allow experts to share insights and predictions about sporting events (eventually other events), and be compensated financially (with tokens) if their predictions are correct. They are well positioned to leverage the new regulatory framework in the United States, as sports gambling has now been widely approved.
Researchably has built an easy to use, personalized AI search engine targeted for individual researchers and clinicians. Individual users can use the search engine to quickly narrow in on research papers and case studies of interest.


Problem that your startup is solving

ThinkCyte has built the world’s first high-throughput flow cytometer, that can be used to both measure and sort cells on the basis of their cell shape and appearance, rather than just size. This is executed via a hardware-based neural network used for image recognition. This tool has potential applications in research as well as clinical environments.