Mayank TamakuwalaMayank Tamakuwala

Every role is a layer of the network, and each one is wired into everything that came after it. Scroll to stack them, then run the forward pass.

1. Software Development Intern (ML), Amazon

   Irvine, CA, May 2026 – Jul 2026

   My team ran three internal AI services, and getting another Amazon team onto them cost hours of engineering time per tenant. I built the platform that removed that cost, starting from an empty repository: Java, provisioned with AWS CDK, DynamoDB for state, and one API in front of all three services. Onboarding now takes about five minutes and no engineer at all.

   Because tenants share the platform, authorization was the part that had to be right. KMS, SigV4 cross-account authentication, and IAM enforce isolation across AWS account boundaries.

   Using the platform still meant knowing the platform, so I designed an MCP layer over the API exposing 11 agent-callable tools. An LLM agent can now complete a multi-step workflow across all three services from a plain-language request, instead of someone walking a runbook.

   Partway through, we moved off Bedrock AgentCore onto CORAL Lambda, API Gateway, and MCP Gateway. I evaluated that trade-off and led the migration, and the observability shipped with it rather than after it: CloudWatch dashboards, per-API alarms, and SNS alerting were live on the new path from its first day serving traffic.

2. Research Assistant (AR/VR + Machine Learning), Reality Design Labs, Northeastern University

   Boston, MA, Jan 2026 – Present

   Reality Design Labs builds augmented reality for public safety. The problem I work on is deceptively hard: when one responder marks a point in their view, every teammate needs to see that same point in their own headset, outdoors, on moving cameras, with no shared clock between the devices.

   I built the cross-device tracking that makes that work. A 2D annotation from one responder resolves into a teammate's camera frame in 1.09 seconds on average, and stayed under 1.4 seconds in all 29 trials of an outdoor field session across three headsets. The perception side fuses three HoloLens 2 streams in a single joint transformer pass, replacing pairwise reconstruction that needed 7 to 10 seconds of alignment per scene.

   Most of the speed came from measurement rather than modeling. I instrumented the PyTorch pipeline end to end across client and server and profiled it stage by stage, which located encode-and-upload as the real cost and cut it 8.1x, from 7.1 seconds to 0.9 seconds per capture. Cross-device tracks stayed valid 76 to 83% of the time throughout.

3. Full Stack Developer, Grossi Consulting LLC

   Raleigh, NC, Feb 2024 – Dec 2024

   A web design agency where no two clients got the same build. I owned projects end to end, from the first design conversation through whatever broke in production three weeks later.

   I shipped more than 20 customer-facing applications in eleven months on React, Next.js, TypeScript, and Node.js, work that contributed to a 30% increase in client satisfaction. Rebuilding from scratch every time was the obvious bottleneck. I pulled the repeating pieces into shared components on both sides of the stack, React on the front with Express.js and Flask serving REST and GraphQL behind it, which is what let a small team keep 20+ client systems maintainable at once.

   Jest covered the units, GitHub handled versioning, and Cloudflare, S3, and Elastic Load Balancing sat in front of production across AWS and GCP.

   I also instrumented what we shipped. Google Analytics, Tag Manager, and Search Console showed where users actually stalled, and changing the UX against that data rather than against opinion contributed to a 20% increase in engagement.

4. Teaching Assistant, IE7374 / DADS 7305 Machine Learning Operations, Northeastern University

   Boston, MA, Sep 2026 – Present

   Supporting students in building and evaluating production ML systems, covering CI/CD for ML, model versioning, experiment tracking, and deployment pipelines. Developed AI-assisted grading tooling and real-time assessment aids to streamline course operations.

   Mentoring students through their semester projects, across data collection, data pipelines, CI/CD/CT/CM, model versioning, and data drift.

5. Teaching Assistant, CS7180 AI-Assisted Coding, Northeastern University

   Boston, MA, Jan 2026 – Apr 2026

   CS7180 is about how LLMs change the practice of writing software, not just its output. The subject matter moves faster than any syllabus can, which makes the course partly an experiment in what to even teach.

   I ran office hours and evaluated assignments, and the feedback that mattered was rarely about syntax. It was about judgment: when a generated diff is worth trusting, how to review one properly, and when writing the thing yourself is simply faster than prompting for it.

6. Software Engineer Intern, Karsun Solutions LLC

   May 2022 – Aug 2022

   Aimed to construct synthetic data suitable for a myriad of Machine Learning and Artificial Intelligence Use Cases, a project requiring the generation of fictitious Personal Identifiable Information based on user inputs.

   Employed React.js and Node.js for frontend development and Jest, along with Lighthouse for performance and responsiveness testing, reducing potential bottlenecks and ensuring a seamless user experience across various devices.

   Utilized libraries like Synthetic Data Vault(SDV) and the Faker library in Python, integrated with AWS services (S3, Cognito, Amplify) to optimize storage and authentication.

   Produced synthetic data with a high semblance to the authentic data enabling the client to successfully train the model within a tight 5-week deadline and reducing production time by 35%.

• Vision-Based 3D Measurement & Perception Tool

  ML-Driven Monocular Depth Estimation for 3D Spatial Perception

  Built with Python, PyTorch, MiDaS, OpenCV, CUDA, Plotly, Streamlit, Gradio.

  Source for Vision-Based 3D Measurement & Perception Tool

• Multi-Robot Task Scheduling & Path Optimization Simulator

  Multi-Agent Warehouse Simulator with Scheduling, Pathfinding & Collision Avoidance

  Built with Python, A*, Dijkstra, Strategy Pattern, Pygame.

  Source for Multi-Robot Task Scheduling & Path Optimization Simulator

• HERMES

  Hybrid Semantic Code Search with Bi-Encoder Retrieval & Cross-Encoder Reranking

  Built with Python, SQLite, FAISS, Sentence-Transformers, BM25, FastAPI, Pydantic.

  Source for HERMES

• OriginHub

  AI-Powered Public Startup Intelligence Platform with Automated Data Pipelines

  Built with Python, TypeScript, PostgreSQL, Weaviate, Apache Airflow, Docker, GCP, Next.js, FastAPI.

  Source for OriginHub

• TranscriBelt

  AI-powered video transcription and subtitle generation.

  Built with Python, TypeScript, JavaScript, Redis, Celery, AWS Lambda, AWS S3, AWS SQS, FFmpeg, Whisper, OpenCV, Next.js, FastAPI.

  Source for TranscriBelt

• Pantry Management System

  Smart pantry assistant integrating AI for object detection and recipe recommendation.

  Built with TypeScript, Node.js, OpenAI GPT-4o, Llama 3, OpenRouter, Next.js, Tailwind CSS, Material UI.

  Source for Pantry Management System