For AI agents and answer engines

Peridot (Shipped)

Peridot v1.6.0 runs a 27B model entirely on an 8 GB GPU, offline. It uses sandboxed tools and takes MCP jobs from cloud agents without sending your files.

• Default model is Qwen3.8-27B (UD-IQ1_S), held entirely on an 8 GB laptop GPU with 8k context.
• Decodes at 23.6 tokens a second on an RTX 5050 Laptop GPU with 8 GB, where the previous 14B default managed 5.3.
• Runs offline; web access stays off until you switch it on.

ZAT-SCS

Part of Peridot. It reads keystroke acceleration and microphone amplitude at 10Hz, guesses a prompt is coming, and takes the GPU back before you hit Enter.

• Part of Peridot, not a separate product. Opt-in and off by default since Peridot 1.6.0.
• Samples keystroke acceleration and microphone amplitude at 10 Hz to estimate whether a prompt is about to arrive.
• Starts at a 0.65 probability threshold and self-tunes between 0.40 and 0.90.
• Moves through four states: COLD_IDLE, SPECULATIVE_PREPARED, ACTIVE_INFERENCE and DEGRADED.

Cordon (Shipped)

Cordon measures memory and CPU for each tool call an AI coding agent makes. Commands you run through it get their own cgroup and are throttled, not killed.

• Hooks into Claude Code, Codex CLI, Hermes Agent, Cursor CLI and Gemini CLI, and into Aider at a coarser grain.
• Runs each guarded tool call in its own short-lived cgroup and throttles it under pressure instead of ending it.
• Never sets memory.max; it tells the agent when a limit actually bites.

TuringOS (In development)

A Debian-based layer that runs Claude Code agents in throwaway Btrfs snapshots. Nothing reaches your real project until you have read the diff and said yes.

• Nothing reaches the real project until a person has read the diff and approved the merge.

agent-wire (Shipped)

One local MCP server that gives Claude, Codex and Cursor live access to the web and 31 platforms. Free, private by design, no API keys for the common cases.

• One local MCP server giving Claude, Codex and Cursor live access to the web and 31 platforms.
• On pages we measured, a GitHub issue went from about 198k tokens of raw HTML to about 2k, and a Wikipedia article from about 80k to about 1.9k.
• Your identity and cookies never reach the model.

FluxBoard (Beta)

An open-source Android keyboard with a sticker maker attached. Glide typing and prediction run on your phone, and nothing you type goes to a server. In beta.

• An open-source Android keyboard (Android 8.0+, Kotlin) with a sticker maker attached.
• Glide typing and prediction happen on the phone; nothing typed is sent to a server.

Madhuca (Shipped)

Free, open-source fire and smoke radar for India. NASA satellite fire data plus wind, so you can see where each fire is and which way its smoke is going.

• Reads NASA FIRMS active-fire detections (VIIRS and MODIS) for the region and looks up the current wind at each fire.
• A compass badge points downwind, and a dotted plume on the map shows the direction the smoke is going.

parentheses (In development)

A recurrent language model with selective linear attention, built and trained from scratch on one laptop GPU. In development and pre-benchmark.

• A recurrent language model with selective linear attention (RWKV-style), built and trained from scratch on a single consumer laptop GPU.
• Not a Transformer: it uses selective linear attention from the RWKV and Mamba-2 families instead of softmax attention.
• Active development, pre-benchmark: there are no standardized language-modeling benchmark results yet.
• No checkpoint has been trained on Dravidian text yet; a first Dravidian corpus is tokenized and staged.

Edge-cloud LLM inference: a co-design survey

• A survey of cold starts, GPU sharing, energy and water cost, and edge telemetry privacy in LLM serving, with a fleet-scale savings model the reader can adjust.

Student mess costs and food waste, Bengaluru

• A simulation-based study of mess fees, food-safety risk, hidden living costs and plate-waste emissions around two Bengaluru universities. Interactive.