I build AI products.
Then put them
to the test.
Previously at Microsoft, Cloud + AI
Product Manager / Software Engineer Intern
A blind benchmark testing whether frontier coding agents follow malicious instructions hidden in repository files, documentation, and setup scripts.
Security benchmarks, useful agents,
and smaller models that reason.
Tested 10 frontier models across 1,950 runs and three safety modes, using fake secrets as tripwires. A Node runtime with a virtual filesystem and shell isolated parallel runs without containers or VMs. Every model executed hidden attacks, with even the safest failing 18% of runs under the default prompt.
Built a platform to compare CPT procedure costs across hospitals, insurers, and locations. A Nemotron vision pipeline extracts codes and charges from bill PDFs and photos. A sandboxed ReAct agent uses schema discovery and bounded tool calls over hospital price data to help users understand bills and compare costs.
Co-authored a framework that stabilizes reinforcement learning with LoftQ initialization and rank-stabilized LoRA, keeping KL divergence low over 500 training steps. Reached 57% pass@1 on AIME24, MATH-500, and AMC23 while using 26% less GPU memory, 66% fewer training examples, and half the compute of standard full-parameter RL fine-tuning.
Built a multi-agent system for patent prior-art research. Strands SDK agents communicate over A2A, with retrieval backed by Bedrock Knowledge Bases. AWS Lambda and S3 support the serverless TypeScript application, while AWS AgentCore hosts the agent runtimes for the research workflow.
Built a student rideshare prototype using Google's Route Optimization API to rank candidate routes by detour cost. Designed the PostgreSQL schema for users, routes, and seat reservations behind the driver and rider flows. Pitched the product and business model at BeaverHacks 2025, winning the Entrepreneurial Challenge.
Built an on-device pipeline with YOLOv11 detection, ByteTrack tracking, and stereo SGBM depth at about 30 FPS with CUDA. ZeroMQ streams object positions to a C++ Steam Audio engine, which renders HRTF spatial cues through bone-conduction headphones. Dockerized the stack to fit within the Jetson Orin Nano's 8 GB of memory.
Reverse-engineered OSU's registration API and built a Chrome extension that emails students when seats open in subscribed classes. A per-minute cron job runs behind a Dockerized Flask REST API and MySQL database, with bcrypt-salted PINs and SMTP alerts. The extension reached 37 active student users and won Best Overall at BeaverHacks.
Built a continuous radio show with a GPT-written DJ covering that day in history every 15 minutes between retro songs. Automated scripting, Google Cloud TTS voicing, music sourcing, and PyDub mixing. Flask on AWS EC2 streams the show to a $10 Goodwill radio rebuilt with an ESP32 and a Three.js listening room.
Built a Chrome extension that compares scraped page keywords with a user's study topics through sentence-transformer embeddings and cosine similarity. Reduced classification error by 75% by averaging each keyword list into one representation, then tuning a weight, error term, and blocking threshold with Nelder-Mead optimization on labeled sites.
Built a Python CLI that combines Reddit posts or GPT-generated stories with narration, synchronized captions, and scraped background footage. The pipeline automates content sourcing, scripting, text-to-speech, and video assembly, then exports ready-to-post MP4s. It uses the OpenAI API, PRAW, gTTS, MoviePy, and PyTube.
Led an eight-person team through Agile sprints, planning architecture, API contracts, and data schemas in PlantUML. Built a React Native and Expo app with a Flask backend on AWS EC2. A BeautifulSoup service provides dining menus and club schedules, with caching for fast loads and JWT authentication for student sign-in.
Built an operating system in C and x86 Assembly with paging, system calls, fork, ELF loading, and preemptive round-robin scheduling through timer interrupts. Implemented user/kernel isolation with protected-mode GDT/IDT setup, trap and page-fault handling, and permissioned page-mapping IPC. Used QEMU and GDB to debug kernel triple faults instruction by instruction.
Built a pipeline that uses AI agents to tailor LaTeX résumé variants to job descriptions. GitHub Actions compiles every variant, publishes PDFs to a rolling release on each push, and cancels stale runs. Reusable agent skills define layout, tone, and content rules, with automated one-page and line-wrap PDF checks.
Built a Spyfall-style party game at BeaverHacks Winter 2023. Players join browser-based lobbies and question each other to identify a hidden spy. A Flask server synchronizes game state for up to 20 players, with a vanilla JavaScript frontend on Replit. Shipped a playable build over one weekend.
Built a Desmos model that links money supply to GDP through linear and composite functions. Students adjust sliders to test how money-supply shocks affect the economy. OSU's Macroeconomics course adopted the model for live lecture demonstrations, reaching more than 200 students per term.
Places I've built things,
broken things, and learned
from people better than me.
Turned an executive request into a shipped planning dashboard through eight stakeholder interviews, requirements, success metrics, and a roadmap. Built an AI agent that turns incident history into runbooks, cutting authoring from days to hours. It won Best Project for AI Innovation among 38 interns, and a second Microsoft organization adopted the framework.
Lead product and engineering for a multi-agent system that researches patent prior art. Own the roadmap and architecture across Strands SDK, AWS AgentCore, A2A, and MCP, working through the tradeoffs between cost, latency, security, and report quality.
Built physics-informed neural networks in PyTorch to model fluid flow and heat transfer inside microchips. Embedded 3D Navier–Stokes and heat equations in the loss, matching analytical profiles within ±5%.
Co-built an advising chatbot with a six-person team for 500+ business freshmen each year. Designed LlamaIndex retrieval around the college’s course catalog and advising documents, then deployed it on AWS with an upload portal professors could use to update its knowledge.
Support 300+ residents with study strategy, time management, and campus resources through one-to-one check-ins. Design and run two academic workshops each term, drawing 50+ attendees each.
Researched the adversarial robustness of implicit neural representation image classifiers. Evaluated PGD attacks and novel defenses across MNIST, CIFAR-10, and CIFAR-100.
Supported a community of 35 residents through conflict resolution, wellbeing check-ins, and connections to campus resources. Organized four community events per term with 20+ attendees each.
Tutored 500+ students in calculus and discrete math through small-group problem solving. Contributed to a TA program that brought the D, F, or withdrawal rate from 33% to 7%. Named Best Teaching Assistant in 2023.
Contributed test-driven development to Dr. Donnelly’s MusicAI Python library and presented the work at OSU’s URSA Spring Symposium.
I'm Ajinkya, pursuing a Master of Science in Computer Science at Oregon State University with a 4.0 GPA, graduating in June 2027. My education also includes an Honors BS in Computer Science with a minor in Economics.
I like figuring out what people need, building something they can use, and testing where it breaks. I lead system design for OSU’s Product Management Club. Away from work, I write songs, play seven instruments, DJ, snowboard, and dance with Kinetic K-pop.
Seven instruments / Snowboarding / Kinetic K-pop
Python · TypeScript · JavaScript · C / C++ · SQL · Bash · x86 Assembly · HTML / CSS · LaTeX
PyTorch · Transformers · vLLM · QLoRA · GRPO · LlamaIndex · RAG · YOLO · OpenCV · PINNs
AWS Bedrock · AgentCore · Lambda · S3 · Strands SDK · MCP · A2A · Vercel AI SDK · OpenRouter
React / Next.js · React Native · Flask · FastAPI · Convex · PostgreSQL · Docker · GitHub Actions
NVIDIA Jetson · CUDA · ZeroMQ · Steam Audio · ESP32 · QEMU · GDB