PRE-FLIGHT

CREW

AMEYA BORKAR

AI/ML · SYSTEMS

MISSION AX-01 · PAD 39_

AMEYA
BORKAR

i build things. some of them even work.

Press and hold — or just scroll.

T+00:00:04

TOWER CLEARED

third year cs at vit pune. two papers, a patent, and about eleven thousand downloads of things i mostly wrote to find out whether they'd work.

TROPOSPHERE · 0–12 km

where it started

wrote a banking system in c — stock market, fraud checks, the lot — just to see if i could. i could. so i wrote a 70,000-line hypervisor microkernel next.

that's more or less the whole pattern. learn a thing, immediately try to build something out of proportion with it, find out what breaks.

  • the bishop's school, pune — icse, 92.1%
  • pioneer public school, pune — class 12, 81.4%
  • vit pune — b.tech cs (ai/ml), third year, cgpa 8.94
  • uc berkeley — summer 2026, cs 160 and cs 161
  • seed infotech — first job, teaching c to thirty-odd students between school and college

weather balloons drift through this layer. pop one.

MAX Q · MAXIMUM DYNAMIC PRESSURE

structural loads peaking. more stress here than anywhere else in the flight.

STRATOSPHERE · WAYPOINTS

flight log

oldest at the bottom, so the list climbs the way the rocket does.

Treasurer, Board of Directors

  • on the board as treasurer. budgets, books, and making sure the service projects have the money to actually happen.

Backend Developer

  • wrote the backend for an agritech company's production site — express, supabase, razorpay, and the rest apis behind a next.js frontend.
  • four weeks. it's live and people actually use it.

AI/ML Lead · VIT Pune

  • run the ai/ml vertical. ten-odd workshops on agentic ai, nlp and vision for around a hundred members.
  • pull in real industry work where i can find it. kissankonnect came in this way.

AI Engineer

  • a chatbot that turns plain english into sql over a proprietary catalogue.
  • most of the actual work was the security layer — prompt injection, anonymisation, making sure it couldn't be talked into showing something it shouldn't.
  • six weeks, shipped to internal users.

AI/ML Developer

  • built and shipped models for predictive analytics, nlp and vision.
  • ran sessions on transformer architectures for about twenty other developers, which taught me more than it taught them.

Computer Vision Head

  • built the vision stack for an autonomous aeromodelling platform — yolov8 detection, tracking, roi — small enough to run on a jetson nano.
  • five of us on it. the team took AIR 1 for overall performance and AIR 1 for design report at SAE 2025, chennai.

100 km

YOU ARE NOW IN SPACE

Kármán line crossed · atmosphere below · sound stops here

MESOSPHERE · PUBLISHED WORK

papers, and one patent

ICAIN 2025 · SPRINGER · FIRST AUTHOR

Adaptive TabNet–β-VAE Fusion Ensemble Closes Uncertainty Gap in Fraud Detection

an ensemble pairing tabnet with a β-vae so the model knows when it doesn't know.

SPRINGER

ICICI 2026 · IN PRESS

EdgeCAAI-Net: Artist-Invariant Network for On-Device Music Genre Classification

the standard benchmarks leak artists across train and test. split by artist instead of by track and macro-F1 drops 0.105 — 18.8% of the score, gone. mine included. gradient-reversal training claws some of it back, in 1.42M parameters.

SOURCE

PATENT FILED

Hypervisor-Based Operating System Isolation

using vt-x / amd-v as the actual isolation boundary for an operating system, instead of ring levels. filed on VeridianOS.

LOW EARTH ORBIT · PAYLOAD MANIFEST

deployed

six payloads on station. click one in orbit — or take it off the manifest.

Problem
os isolation leans on ring levels and page tables sharing one address space. large attack surface.
Approach
every component runs as a hardware-isolated domain inside its own nested page-table tree. own smp bring-up, ahci/nvme drivers, ext2, iommu, and a tcp/ip stack written from scratch.
Outcome
70,788 lines. boots end to end on kaby lake, tiger lake and zen 5. patent filed. i'm still processing this one.

GITHUB

Problem
distributed rate limiters quietly overshoot across a fleet, and say nothing about the two axes that actually cost money: concurrency and llm tokens.
Approach
GALE holds global overshoot to a bound independent of fleet size, machine-checked in tla+. TALE meters llm output tokens you only find out about as they stream.
Outcome
169 ns in-process check, zero runtime dependencies, bit-identical across six backends. ~9,700 downloads. roman voloboev, who wrote rate-limiter-flexible, said nice things about it.

LIVESOURCE

Problem
a vector database with a search bar isn't memory. no sense of recency, picks a side silently when sources disagree, can't tell you what was true last tuesday.
Approach
coarse-to-fine retrieval over a decaying hierarchy. contradictions surface together instead of one being chosen. queries answerable as of any point in time.
Outcome
87.75% on LongMemEval-S (n=500) with an open-weight actor. on pypi as engrampy.

SOURCE

Problem
map apis charge per call and restrict what you're allowed to keep.
Approach
key-free engine on openstreetmap data: nearby ranking, isochrones, walkability, amenity gaps. the errand planner solves a generalized tsp with exact held-karp dp over bitmask states.
Outcome
three npm packages at v1.0 — a library, a cli, and an mcp server so agents can use it as tools.

SOURCE

Problem
wanted to actually use the data structures course instead of just passing it.
Approach
an emergency dispatch sim of a grid city where every live number on screen is computed by a different structure. fibonacci heap, suffix array, bk-tree, segment tree, quadtree, about two dozen in all.
Outcome
the whole ds syllabus, cosplaying as a 911 dispatcher. c11, no dependencies.

SOURCE

Problem
there wasn't one. i wanted to know if i could write a bank in c.
Approach
15,000 lines of console banking. concurrent sessions, a live stock-market simulator with quantitative analytics, a four-heuristic fraud engine, real-time messaging, pin and 2fa.
Outcome
i could. that's the entire outcome. it's also why VeridianOS exists — after this, an operating system felt like a reasonable next step.

GITHUB

more on github — other kernels, ml research that didn't work, finance tools, and web things without a build step.

APOAPSIS

currently

still writing kernels in my spare time. it boots, which surprises me more than it should.

mostly i'm chasing the same thing i always am: find something i don't understand, build the version of it that's slightly too ambitious, and see which part breaks first. two years of that and i haven't got bored of it yet.

  • 5am lifts. by choice. i don't know why either.
  • making machines smarter than me, which is mildly concerning.
  • more papers in the drafts folder than out of it.

that dot is Earth

TRANSMISSION · DISH LOCKED

say hello

open to interesting builds, research, and conversations. i answer everything.

DOWNLINK ameyaborkar17@gmail.com

GITHUB145.800 LINKEDIN437.525 EMAIL121.500 DOSSIERPDF

Technical skills

Languages — Python, C, C++, Java, Go, TypeScript, JavaScript, SQL, x86-64 Assembly

AI/ML — PyTorch, TensorFlow, scikit-learn, Hugging Face, OpenCV, YOLOv8, XGBoost, transformers, CNNs, RNNs/LSTMs, VAEs, adversarial and contrastive training

LLMs & agents — LLM fine-tuning, prompt engineering, agentic AI, NL-to-SQL, ReAct agents, MCP, embeddings, RAG

Systems — Linux, Docker, Redis, Postgres, gRPC, TLA+, distributed systems, multithreading, hypervisors (VT-x/AMD-V), IOMMU, TCP/IP

Web & backend — Next.js, React, Node.js, Express, FastAPI, Flask, Astro, Supabase, REST APIs, MySQL

Spoken — English (professional), Marathi (native), Hindi (native)