Mehran Mozaffari

Do not debug a manipulation policy on hobby servos. Internal gear slop turns closed-loop contact into an open-loop guessing game.

ooosplat is a 3D Gaussian splatting tool that reconstructs navigable scenes directly from casual phone footage. The page does not state compute budgets or training runtimes, and automated splatting still degrades on reflective surfaces, motion blur, and low-parallax pans. What strikes me is treating fast scene reconstruction as plumbing for embodied simulation environments rather than a standalone graphics demo. github.com/ooolabdev/ooospla…

browse is an MCP server and Chromium extension that lets coding agents control tabs inside an existing browser profile. The part not on the project page is pausing at 2FA and login barriers for manual handoffs while capturing HAR logs alongside screen recordings. What strikes me is treating human intervention as an explicit protocol event rather than pretending agents can bypass anti-bot defenses unattended. opencode.ai/product/browse

If an avatar cannot instantiate in a single forward pass, you do not have a conversational product. Nobody waits on test-time optimization.

Open weights solve dialect coverage, but interactive agents die on turn latency long before regional nuance matters.

Coding harnesses like Cursor and Claude Code run subagents in isolated contexts instead of shared sessions. Compressing context into dispatch prompts and summary returns makes the delegation boundary the most fragile seam. piazr.github.io/applied-ai/a…

Continuous background extraction will probably give way to query-time search. Pre-building entity graphs on every turn burns tokens on questions nobody asks.

json-render is a library that converts streaming JSON from an LLM into native Slack and Teams message payloads. The part not on the project page is throttling updates to a 500 ms cadence to survive platform edit limits, and silently downgrading over-quota layouts unless you force a model self-repair loop. What strikes me is treating generative UI as a typed intermediate representation rather than ad-hoc tool calls, letting the client absorb platform constraints instead of the prompt. vercel.com/changelog/json-re…

How do you escape the choice between massive KV-cache footprints and brittle top-k retrieval? Probably treating memory as a continuous modality rather than a search index.

Obsessing over geometric fidelity in web viewports is strange. A browser tab crashes on GPU memory long before polygon counts ever matter.

Piling compiler traces into an agent prompt is an odd reflex. Terminal stdout is transient state, yet we preserve it like dialogue history.

media-tooling is a CLI toolkit that gives coding agents deterministic shell commands to cut and render video. The part not on the project page is collapsing video frames and waveforms at cut points into single composite images so vision models can verify edits in one token budget. What strikes me is treating video editing as an inspectable compiler pass rather than waiting for multimodal models to output finished files. github.com/kumanday/media-to…

Generative UI agents output flat pixels instead of compiling declarative state trees. Raster outputs discard semantic structure, forcing every subsequent iteration to re-synthesize the entire layout from scratch. piazr.github.io/applied-ai/a…

How do you stop an autonomous coding agent from drifting into loops? Probably amnesic process resets, letting git state replace conversational context.

security-audit-skill is a harness that directs coding agents through multi-phase vulnerability sweeps across a codebase. The part not on the project page is decoupling discovery from verification so the agent that spots a vulnerability cannot validate it, leaving a fresh isolated agent to disprove it. What strikes me is treating an unresolved fact as an explicit status rather than forcing the model to guess a verdict. github.com/cloudflare/securi…

Local model runners like Ollama run as privileged host daemons instead of isolated sandboxes. Giving a runner ambient access to shell history and private keys makes native daemons an indefensible stopgap. piazr.github.io/applied-ai/a…

Serializing an order book so agents can debate it feels strange. You swap deterministic matrix math for committee role-play on structured time series.

Web frontends will probably stop downloading dense .glb files. Offline engines still need frozen meshes, but compiling visuals straight into Three.js code keeps interactive assets mutable.

Agent frameworks like LangGraph freeze execution scaffolding in host code instead of letting runtimes self-adapt. Letting models generate dynamic tools does not help when the outer retry loops and middleware remain completely rigid. piazr.github.io/applied-ai/a…

esp32-iphone-agent is a hardware harness and iOS app that lets vision models observe an iPhone screen over AirPlay and drive it with USB mouse emulation. The part not on the project page is visually closing the input loop by homing the cursor in a corner and wiggling it to measure relative displacement. What strikes me is that dodging mobile sandboxes with external hardware just trades permissions for visual servoing heuristics, though the page does not state step latency. github.com/onehumanbeing/esp…