Oct 2021 – Jan 2025 · 3 yrs 3 mos

• Senior-most engineer at Amazon Delhi NCR amongst 700+ engineers; Amongst senior-most at Amazon India; Architect for over 10 projects worldwide; consultant for many others; Featured by Amazon on its global LinkedIn channel [D1]; Team spread across India, USA and Australia
• Ownership: Online identity verification solutions for Amazon E-commerce worldwide, except India, for fraud and abuse prevention and regulatory compliance, using both AI automation and human workforce
• Projects Cairo/Suez/Giza/Handshake: Artificial Intelligence for ID document automation for identity verification, one of the most innovative projects from Amazon India (Details under AWS section below)
• Wrote a VP-level proposal (In-Person Verification, IPV [D2]) in July ’19 resulting in creation of four new projects, led the technical architecture, resulting in overwhelmingly positive business impact; The VP wrote 2+ years later in Nov. ’21, “And I talk about IPV a lot - it was a different way of thinking about this {…} You should feel proud!”
• Tech Lead for Alexandria, Amazon-wide secure storage solution for critical documents. Aligned multiple Amazon orgs towards simplification and project deduplication resulting in ten+ person-years of projected savings.
• Delivered many technical and semi-technical talks within the company as well as externally. See my featured posts for links to slides and recordings.
• Note: My work at Amazon CTPS and AWS was largely in continuity. The projects mentioned cross the timelines for when I was with CTPS vs. AWS.
• [D1] https://www.linkedin.com/posts/amazon_techjobs-insideamazon-amazonians-activity-6812952277926653952-nfkm
• [D2] https://www.geekwire.com/2020/amazon-tests-using-live-video-calls-vet-aspiring-third-party-sellers-seeking-weed-fraudsters/

Jul 2020 – Oct 2021 · 1 yr 3 mos

• Projects Cairo/Suez/Giza/Handshake: Artificial Intelligence for ID document automation for identity verification; Lead technical architect since 2016.
• Amazon CEO for Worldwide Consumer [T1], Dave Clark, praised the project as one of the inspiring projects from Amazon India, “… team developed a vision-based information extraction capability used to automate identity verification, helping streamline new seller onboarding experience worldwide.” (May ’22) [T2]. A Sr. SDM called it “one of the most challenging products I’ve seen in this org (both technically and product-wise).”
• Best known automated ID document forgery detection performance for US ID documents per the internal benchmarks and limited third-party vendor evaluations (as of ~2024)
• Originally developed for internal use, AWS Textract Analyze ID was launched from this in Dec. ’21 [T3-5] by a sister team. Included by VentureBeat in the top 12 items in AI/ML from re:Invent 2021 [T6]. AWS Rekognition Liveness later launched from this in April ’23 [T7-8].
• [T1] https://ir.aboutamazon.com/officers-and-directors/person-details/default.aspx?ItemId=be07815d-9c56-4984-82a1-2cf131ddf0b2
• [T2] https://www.business-standard.com/article/companies/inspired-by-indian-innovation-says-senior-amazon-executive-dave-clark-122051801689_1.html
• [T3] https://docs.aws.amazon.com/textract/latest/dg/how-it-works-identity.html
• [T4] https://aws.amazon.com/blogs/machine-learning/announcing-support-for-extracting-data-from-identity-documents-using-amazon-textract/
• [T5] https://www.linkedin.com/posts/alokgovil_announcing-support-for-extracting-data-from-activity-6872070419768332288-vppC
• [T6] https://venturebeat.com/2021/12/03/top-12-ai-and-machine-learning-announcements-at-aws-reinvent-2021/
• [T7] https://aws.amazon.com/rekognition/face-liveness/
• [T8] https://aws.amazon.com/about-aws/whats-new/2023/04/amazon-rekognition-face-liveness-deter-fraud-facial-verification/

Aug 2016 – Jul 2020 · 3 yrs 11 mos

• (My work at Amazon CTPS and AWS was largely in continuity. The projects mentioned under CTPS and AWS cross the timelines and cover this period as well.)

Aug 2015 – Aug 2016 · 1 yr

• Amazon hired me as amongst the senior-most engineers in the Alexa NLP/NLU org of about 85 people based in part on my ‘Thought t;’ experience (described later below) where I had dreams to build something like ChatGPT. I joined Amazon thinking that I would have the opportunity for the same in the Alexa team, however, I did not quite get the environment I needed (even though my manager/Director was by himself very supportive) in part because of prioritization of the existing projects at Amazon and in part due to a personal challenge I was facing behind the scenes.
• During my time in the Alexa team, I made an in-depth proposal for holistic integration of Evi and Alexa, unaccomplished for 3 years of Evi’s acquisition (Evi [E1] was a Question-Answering AI that Amazon had acquired). I pushed for research work in Semantic Parsing and aligned the extended organization on a proposal to streamline integration of new knowledge categories into Alexa Knowledge Graphs. I also defined a multi-year vision for Alexa expansion into a new business domain (NLIDBs [E2]). While this proposal was considered too risky at the time and was not funded, AWS QuickSight Q later launched in Dec. ’20 [E3]. I also performed technology evaluation for Graphiq (startup) acquisition for integration with Alexa [E4].
• In about a year, I was drawn towards a very interesting research project in Computer Vision for ID document automation, based out of India. Taking this project was also a good direction for me to pursue for family reasons.
• [E1] https://en.wikipedia.org/wiki/Evi_(software)
• [E2] https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=00e26133551f152f4c8913c5442238d30a63ef79
• [E3] https://aws.amazon.com/about-aws/whats-new/2020/12/introducing-amazon-quicksight-q/
• [E4] http://www.latimes.com/business/technology/la-fi-tn-graphiq-amazon-20170719-story.html

Jan 2014 – Aug 2015 · 1 yr 7 mos · Santa Clara, CA (USA)

• Project: Next-gen Computer Vision platform for Edge AI, reimagined from hardware & algorithms to applications
• Paved the way for a novel sensor module, Glance, with >500x reduction in power. The seminal work has been covered by several news/blogs including MIT Technology Review [G1-6].
• Led a multidisciplinary team with computer vision, machine learning experts, software developers and analog/digital circuit designers. Grew the project from 1 person (self) to >10 fulltime engineers. Worked with marketing leads for product definition and meetings with Tier 1 vendors and customers.
• Developed system emulator in C/C++ for proof-of-concept and demos
• Primary inventor on eight filed patent applications, including one receiving 'Gold' rating (top 0.5%)
• Led to the Tiny ML industry trend, co-founded by my manager at Qualcomm
• Launched in Dec. ’21 as a part of Snapdragon 8 Gen 1 [G7-8]
• [G1] https://www.technologyreview.com/s/603964/qualcomm-wants-your-smartphone-to-have-energy-efficient-eyes/
• [G2] Kristof Beets at Embedded Vision Summit, "... Qualcomm’s always-on vision solution ... takes efficiency to the extreme. It is ultra-low-power and fully integrated ...", https://www.imgtec.com/blog/embedded-vision-summit-2017-thoughts-from-the-floor/
• [G3] https://image-sensors-world.blogspot.com/2017/05/more-info-on-qualcomm-always-on-vision.html
• [G4] https://www.qualcomm.com/invention/research/projects/computer-vision/always-on
• [G5] https://www.embedded-vision.com/summit/always-vision-becomes-reality
• [G6] The Tiny ML book cited the work on its first page, “pioneering projects … even in the vision world (Qualcomm’s Glance camera module).” https://www.amazon.com/TinyML-Learning-TensorFlow-Ultra-Low-Power-Microcontrollers/dp/1492052043/
• [G7] https://www.qualcomm.com/news/onq/2021/11/30/snapdragon-8-gen-1-mobile-platform-world-firsts
• [G8] https://www.linkedin.com/posts/evgeni-gousev-2832858_fyiand-if-you-happen-to-be-in-the-bay-area-activity-6262810102156320768-GUQb

Dec 2012 – Dec 2013 · 1 yr · Fremont, CA (USA)

• I had interest in AI from early childhood when I saw computer play chess against my grandfather, and in ‘teaching’ English to machines from tenth grade when I had built a prototype compiler [F1-2]. Purely via self-learning, by 2013, I had world-class expertise in Natural Language Processing (NLP). Later in 2014, I delivered a large-scale seminar on the same [F3].
• In Dec. 2012, I dropped my fulltime job with a dream to make something like ChatGPT. I codenamed it ‘Thought t;’ where ‘t’ would represent a thought inside a computer just as ‘int i;’ is for an integer. The work was self-financed while a friend moonlighted.
• We attempted making deep neural networks and training these with text from the entire English Wikipedia [F4-5].
• To continue financing the above, in Jan. 2014, I joined contract work at Qualcomm in the area of computer vision. The new project was so interesting in itself that my above work went on a pause. Later, Amazon hired me at the senior-most engineer in the Alexa org based on my ‘Thought t;’ experience.
• While I did not accomplish the said vision (and have no evidence to hint I could have), I started sharing my thoughts with other engineers and entrepreneurs. When ChatGPT and Gen AI came out, a Principal Engineer wrote, “Hey Alok, I have been telling everyone about you, your predictions and how you were correct!! Chatgpt is insane!!”, while a college friend wrote “Whatever you predicted that time {2017} is happening right now .. 🙏”
• [F1] https://www.linkedin.com/feed/update/urn:li:activity:6844576103705649152/
• [F2] https://alokgovil.com/wp-content/uploads/interview-by-qeemle/Alok-Govil-AI-Could-Augment-Our-Intelligence-Qeemle.pdf
• [F3] http://www.siliconvalley-codecamp.com/Session/2014/introduction-to-natural-language-processing-nlp
• [F4] Resources used: English Wikipedia dumps, Stanford CoreNLP Java library, Neo4j Graph DB, Gremlin
• [F5] Processing English Wikipedia was itself a stretch. Today, LLMs are commonly trained on Internet-scale texts.

Oct 2008 – Nov 2012 · 4 yrs 1 mo

• Nearly when I started at Qualcomm in 2005, I was a proponent of a new technology architecture that was several times more promising, but also correspondingly risky. In 2008, I was made the project lead for R&D into this architecture with a team of 1.5 people including myself. Over a period of four years, I took this project through proof-of-concept to adoption across the entire Qualcomm division, making this the biggest project I have led till date.
• These four years involved performing end-to-end system design from device physics to software and to user-experience, project management and recruiting talent, building consensus across teams on multiple critical issues, delivering quarterly reviews to Qualcomm “C-level” executives, division-wide seminars and courses, motivating and mentoring, and also helping establish relationships with potential tier 1 partners in Asia.
• My work also continuously involved developing system models in Matlab and Mathematica to evaluate the full architecture / design space over performance, costs, time-to-market, and risks, and optimizing the same using genetic algorithms. Given the unique nature of the technology, I also invented [M1] many image and video processing algorithms and implemented them in C++ and Matlab.
• Over the time, I amassed both breadth and depth into multiple technical disciplines as well as project management, and have worked with many world-class subject matter experts. A lot many have considered me to be an expert in their respective fields and in a few cases without realizing that I was doing the same across many other fields.
• I left Qualcomm in late 2012 with dreams to do something still bigger.
• [M1] Six filed patent applications

Apr 2005 – Oct 2008 · 3 yrs 6 mos

• My key reason to join Qualcomm was the multidisciplinary nature of the project, spanning both software and hardware. The excellent work culture allowed for initiative and autonomy, making many of my projects self-initiated.
• My first project involved developing software-development tools for a custom application-specific processor with about 25 instructions that I also architected. I developed an IDE for it in C# including a basic debugger. Instead of developing a full programming language and compiler, I came up with a technique to reuse Microsoft’s C# compiler to compile C# for this processor. The paper I published on this at Qualcomm tech forum reached in the top 10%. I also helped with circuit design, Verilog coding, and the embedded C program.
• A second major project required self-calibration of display modules. I prototyped the analog back-end and developed a behavioral model of the full system in C#. The system was then implemented including embedded software, FPGA board and analog IC design. The latter two were developed by my team based on architecture I defined.
• While working on the above, I was making pixel device measurements that were unexplained. I ended up developing industry’s first analytical model to explain MEMS response-time under damping, and delivered division-wide seminars on it.
• Another project is worth a mention as it went all the way from skepticism to deployment across the division. In spite of fabrication of a lot of display panels, the amount of available test data was still limited. I proposed measuring every pixel using high-resolution cameras and developed this system using C and Matlab. This involved taking thousands of pictures of each panel subjected to varying test waveforms and using image processing to locate and characterize the pixels.
• Overall, I learned a lot on both business and technology fronts, worked with many excellent people, and later led a major research program starting in 2008 (see above).

Feb 2001 – Apr 2005 · 4 yrs 2 mos

• While I had many job offers coming out of USC, I joined Philips Research because the place was full of smart and accomplished researchers (nearly all PhDs). My key focus here was circuit design and software development for research in electronic displays. I went though complete design life-cycle including analog and digital IC (integrated circuit) design, layout/synthesis, software development, PCB design, lab measurements and customer interaction.
• My most challenging project here was designing a sub-femtoFarad-precision capacitance measurement circuit for automated measurement of every pixel of a Full HD LCD device. I could never test this circuit though as Philips changed its research direction [L1] shortly after the IC tape-out.
• I also designed advanced analog circuits for medical ultrasound transmitters exceeding the state-of-the-art by a factor of five in performance and power requirements. The project was so challenging that Philips Medical would not let us publish even the target specifications for the circuit, let alone its design or performance measurements.
• Another awesome project here involved laser-beams and software! We used an LCD micro-display to make an electronic hologram. I developed the software and algorithm in C++ to “compute” holograms from simple 3D input scenes. The project was adopted by a research team in Europe for further development.
• A deep influence here was Dr. Remus Albu [L2], who I consider to be as close to a world-class physicist as I ever hope to meet. He turned me from a mere student of science to a scientist and into the man I am now.
• [L1] http://www.eetimes.com/document.asp?doc_id=1151508
• [L2] http://www.linkedin.com/pub/remus-albu/3/a59/a14

Jan 1999 – Jan 2000 · 1 yr

• Among many other research activities, IMSC was advancing Augmented Reality (AR) using vision-based camera pose [C1] tracking. Live video-stream from a camera mounted on a virtual reality headset (like HTC Vive) was processed in real-time to find camera pose, virtual objects were superimposed and the video stream sent back to the headset.
• Here I wrote a C++/MFC-based application for development of image feature tracking algorithms, developed in record-times (and earning me this research assistant position :-)). I later developed an Augmented-Reality Golf game using the head-mounted display setup where you could see a virtual golf course right within the lab, receiving about 28 citations for the technical paper [C2, C3] published.
• While developing a technology like Augmented Reality, so easy to describe though much harder to implement at least at that time, we were all swimming in ideas about its applications and new research directions. Just one of the things I envisioned then were hardware-accelerated physics engines which were first seen in the market about four years later (PhysX [C4]).
• [C1] http://en.wikipedia.org/wiki/Six_degrees_of_freedom
• [C2] http://graphics.usc.edu/cgit/publications/papers/ACMM2000ARGolf.pdf
• [C3] Citations: https://scholar.google.co.in/scholar?gws_rd=cr&um=1&ie=UTF-8&lr&cites=16091752000003864483
• [C4] http://en.wikipedia.org/wiki/PhysX

Jan 1998 – Jan 1999 · 1 yr

• River Run was a decade ahead of the industry in mobile revolution witnessed around iPhone coming out. Their product lines included complete mobile solutions including mobile apps, server back-end, deployment solutions (think App Stores), and even app-development tools that our team built.
• We developed a GUI-based software in C++/MFC that allowed generation of (source code of) fully functional mobile apps, including the app GUI, navigation and server connectivity. Augmenting this was a database administration tool that allowed complete control of enterprise data with provably minimal user input, built on the top of Lotus Notes, an early NoSQL database [R1].
• While I was gradually getting famous for developing thorough and bug-free code (I received two night-in-town awards in recognition of this), I couldn’t sleep well for two nights when a memory leak was found in my code. A deep influence was a teammate, Viraj [R2], who could write a few thousand lines of C++ code without a single “compilation” error!
• The company unfortunately, I felt, itself failed to anticipate the mobile revolution coming that brought with it a wave of competitors, both big and small.
• [R1] http://www.dbta.com/Articles/Columns/Notes-on-NoSQL/NoSQL-and-Document-Oriented-Databases-72035.aspx
• [R2] http://in.linkedin.com/in/viraj

Jan 1998 – Jan 1998 · 0 mo

• The Department of Electronics had every integrated circuit (IC) design software we had known. My friend Sushil [S1] and I thereby felt very lucky to get an opportunity to design a circuit that the two of us had conceived earlier, but were finding hard to develop further without the tools. This opportunity ultimately lead to the publication of our first conference paper [S2]. I later openly received a few “plagiarism” requests from junior students to copy our work, which I gracefully denied, of course.
• [S1] http://www.ece.lehigh.edu/index.php?page=sushil-kumar
• [S2] http://ieeexplore.ieee.org/xpl/articleDetails.jsp?tp=&arnumber=780670