Skip to content
View aadityad12's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report aadityad12

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aadityad12/README.md

Hi, I'm Aaditya 👋

Computer Engineering student at San Jose State University (Dec 2027) building AI systems down to the hardware: on-device NPU inference, custom BLE wire protocols, and LLM agents and evals. I ship end-to-end deployments, not demo notebooks.

Currently: seeking internships in AI/ML engineering and edge AI / embedded systems.

What I build

🧿 GazeBoard · Lets ALS patients speak using only their eyes, on an off-the-shelf phone. Runs MediaPipe FaceMesh (478 landmarks) on the Qualcomm Hexagon NPU via the LiteRT CompiledModel API at ~8ms inference, fully offline, with a custom 4-point affine calibration engine. Kotlin · LiteRT · CameraX

🪗 Accordion · Reversible context management for LLM coding agents: the whole context window as foldable blocks under a token budget. I built the-conductor-v2, a three-stage relevance pipeline (keyword → bi-encoder cosine → cross-encoder rerank) with self-calibrating fold targets, plus the live conductor activity dashboard. Won The Token Company's sponsor prize at the UC Berkeley AI Hackathon 2026. TypeScript · Node.js · SvelteKit

🌡️ Temper · Evaluates the environment around an LLM (system prompt, tools, skill files) instead of the model itself: scores a harness against a bare-model baseline across six dimensions, generates targeted patches, and re-evaluates to confirm they worked. I built the core eval engine end to end. Python · FastAPI · Gemini

🚒 Clear Dispatch · Four LLM agents behind a 911 dispatcher during wildfire call surges: triage, unit routing, and voice briefings, with the human always in control. Real-time WebSocket pipeline, zero polling. Python · FastAPI · React · Claude

📡 Echo · Emergency alerts that hop phone to phone with no internet or cell service. Custom GATT chunked-transfer protocol with native Kotlin and Swift implementations, gossip mesh with SHA-1 deduplication, Raspberry Pi relay nodes, and offline translation into 23 languages. Flutter · Kotlin · Swift · BLE

Stack

Core: Python · C++ · Kotlin · TypeScript · x86 Assembly AI/ML: LiteRT/TFLite · MediaPipe · scikit-learn · LLM agents & evals Systems: BLE/GATT protocol design · on-device inference · FastAPI · WebSockets

Reach me

📧 aaditya.d.desai@gmail.com · 💼 LinkedIn

Pinned Loading

  1. Echo Echo Public

    H4H, 2026 - V3

    Dart 4

  2. Trackers Trackers Public

    All in one Tracker app

    Kotlin 1

  3. ASM-79/DA_HACKS_3.5 ASM-79/DA_HACKS_3.5 Public

    JavaScript 2

  4. hw528/sfhack hw528/sfhack Public

    Python

  5. Clear-Dispatch Clear-Dispatch Public

    TypeScript 1

  6. GazeBoard GazeBoard Public

    Kotlin