Applied AI engineer. I build the layer underneath the models:
gateways, agent runtimes, and inference that runs on your own machine.

himansh.in · maya.himansh.in · Hugging Face · LinkedIn · X

npm i -g maya-os && maya

One command boots an OpenAI- and Anthropic-compatible gateway that routes a single request surface across cloud providers and local GGUF models (llama.cpp / Metal), with RAM guardrails that degrade instead of killing your Mac.

On top of it: agent teams running on real PTYs. Claude Code, Codex and maya code driven as genuine terminal processes, coordinated by an orchestrator with a persistent loop agent, a message bus, and SQLite-backed shared contracts. Eight packages, MIT, shipped.

maya code is a provider-agnostic agent loop over the Anthropic + OpenAI SDKs with a full-screen Ink TUI, skills, MCP, sub-agents, workflows and mid-turn steering of a running agent.

Models from scratch. F75, a ~100k-parameter LM trained end-to-end on one RTX 5060 Ti to study next-token behaviour under extreme capacity limits. LlamaVision, a 6.3M-parameter projector aligning vision features into a frozen Llama-3.3-1B embedding space.

Systems, low enough to hurt. H++, a statically-typed language whose only primitive is the bit (compiler + runtime in C and Assembly). An 8085 emulator with full 64KB addressing and step debugging. NetGuard, a no-root Android firewall that intercepts DNS through a local VpnService tunnel.

Infra for my own work. boxctl-deck, GPU workstation fleet control over SSH with live telemetry.

Applied AI Engineer @ Colate. Built the multi-provider LLM gateway behind Cana: one shared request/event representation translating OpenAI, Anthropic and Gemini formats, with streaming, tool calls and multimodal inputs. Added prompt caching and token accounting: 6.05B cached input tokens, 53% input-token reuse, 35% lower inference latency. Also built its multi-agent orchestration (worker, reviewer and lead roles, dependency-aware delegation, inter-agent messaging) and persistent AWS/Kubernetes runtimes.

Research Intern @ MISN Lab, IIT Delhi. Agentic systems for recommendation under Dr. Sandeep Kumar: retrieval, planning and tool use, plus the evaluation methodology to tell whether any of it actually works.

HIVE @ IIT Delhi 2024 · FinAThon @ IIT Delhi 2023 · Top 10 worldwide @ IIT Bombay TechFest 2023 · 3rd @ DEXTERIX 4.0

B.Tech CSE (AI & ML), IILM University · 2023-2027 · Noida, India