📫 adityaaguha@gmail.com | 📞 +91 7304748055
- Building the scanner module for MacBenchForge, a macOS Apple Silicon GPU benchmarking tool
- Final-year B.Tech CSE (AI & ML), graduating 2026
- Job hunting for full-time AI Engineer / Applied AI Developer roles (Pune, remote)
Final-year AI & ML engineer with hands-on research at DRDO on reinforcement learning for quadruped robotics, and a track record of shipping local-first AI systems end to end, agentic assistants, LLM inference servers, and custom benchmarking tools built in C++ and Python. I care about privacy-focused, self-hosted AI infrastructure and getting real performance out of consumer hardware. Outside of core dev work, I run a freelance computer consultancy and create AI/dev-focused content for Instagram and YouTube.
Defence Research and Development Organisation (DRDO), Pune Robotics & Machine Learning Research Intern
Project: Reinforcement Learning Based Quadruped Handstand and Footstand using MuJoCo and JAX
- Developed reinforcement learning control strategies for quadruped robotic balance and posture stabilization
- Worked with the MuJoCo physics simulation environment for robotics modeling
- Implemented and analyzed control policies using JAX-based reinforcement learning, including PPO with curriculum learning
- Contributed to simulation-driven learning and control optimization research
Freelance Computer Consultancy
- Independent consulting on hardware builds, benchmarking, and system setup for individual clients
B.Tech, Computer Science Engineering (AI & ML) Bharati Vidyapeeth Deemed University, DET, Navi Mumbai — CGPA ~8.5, Class of 2026
Local-first ReAct agentic AI assistant. FastAPI backend with SSE streaming, llama.cpp inference (Qwen3-8B), Serper.dev web search, and a Perplexity-style vanilla JS frontend.
Local image generation stack: FastAPI + llama.cpp (Gemma, Metal GPU) + ComfyUI running headless on a MacBook Pro M1 Pro, with one-command install and launch scripts.
Open-source C++ GPU/CPU benchmarking tool for Windows, built and tuned against an RTX 4050. (add repo link)
Companion benchmarking tool for macOS Apple Silicon, using IOKit for GPU detection. (add repo link)
Automated road-littering detection and e-Challan system using YOLOv8, PaddleOCR, and FastAPI. In progress. (add repo link)
Final-year project: multimodal AI proctoring system combining MediaPipe face/gaze tracking, YOLOv8, and CNN-based audio analysis for exam integrity monitoring. (add repo link)
More Projects
NeuroCourier / GuhaGPT — Multimodal AI agent on Telegram and Discord, Ollama backend, supporting text and image input for privacy-focused interaction. (add repo link)
MeetingMind — Local-first meeting transcription and RAG system: pyannote diarization, faster-whisper, ChromaDB, Ollama, FastAPI + WebSocket. (add repo link)
VisionSense — Real-time scene description combining YOLOv8 object detection with Qwen2.5-VL-3B. (add repo link)
NutriLens — FastAPI + Gemini Vision Telegram bot for food label analysis. (add repo link)
CortexCLI — Textual-based multi-provider chat CLI, with a planned agentic-tool-calling upgrade (GuhaCLI). (add repo link)
n8n Bank Statement Analyser — Local workflow: PDF bank statements parsed and analyzed by AI, results delivered via Telegram, containerized with Docker.
YOLOv8 Vehicle Detection — Custom-trained 6-class vehicle detection model built on a Roboflow dataset. (add repo link)
MiniZIP++ — Custom C++ archiver with a Huffman coding compression layer. (add repo link)
Local LLM Server — FastAPI backend serving locally hosted LLMs via structured APIs, used across several of the projects above. (add repo link)
Encrypted NAS — Fully encrypted, self-hosted network-attached storage with LUKS + Samba, multi-user access across platforms.
TripMind — FastAPI + React + Gemini 2.5 Flash travel planner MVP. (add repo link)
BhashaMitra — Gamified Indian language learning platform, built for a hackathon. (add repo link)
Celestique — Salon platform startup concept.
IVA — Donor-NGO bridge startup concept.
I run and benchmark a small fleet of machines, and treat hardware tuning as seriously as the software on top of it:
- Lenovo LOQ (RTX 4050) — Fedora 44 with a full CUDA + llama.cpp rebuild, ~103 tok/s on Gemma 4 E2B
- MacBook Pro M1 Pro (14", 16GB) — Metal-accelerated inference benchmarking, ComfyUI, PixelStudio Pro host
- ThinkCentre M91p — Repurposed as an OpenMediaVault NAS (Debian 13), SMB shares, CPU-only llama.cpp benchmarking on Sandy Bridge
- ASUS Vivobook OLED 15 (Ryzen 5 7520U) — Cross-platform llama.cpp benchmarking (Vulkan/CPU)
All machines are tied together over Tailscale for remote SSH access, and I regularly run cross-device llama.cpp benchmarks comparing CUDA, Metal, and CPU-only inference paths.
Languages: Python, C++, JavaScript
AI / ML: PyTorch, OpenCV, YOLO, MediaPipe, Reinforcement Learning (PPO, Curriculum RL)
Robotics & Simulation: MuJoCo, JAX, Gymnasium
LLM Infrastructure: llama.cpp, Ollama, ComfyUI, GPU inference (CUDA, Metal, Vulkan), FastAPI, SSE streaming, RAG (ChromaDB)
Computer Vision: YOLOv8, MediaPipe, PaddleOCR, Qwen2.5-VL
Systems: Linux (Fedora, Debian, Ubuntu), self-hosted NAS, Tailscale, Docker, computer architecture and hardware benchmarking
I post AI and developer-focused content on Instagram and YouTube under @adityaguha_, covering local LLM builds, benchmarking, and practical AI engineering.
- GDSC Chapter Lead (2023-24)
- PR Executive, BVDU DET
- Campus Executive, GeeksforGeeks
📫 adityaaguha@gmail.com | 📞 +91 7304748055 | adityaguha.tech