I maintain AERPAW/aerpawlib — the Python API experimenters use to fly LAM6/SAM4 UAVs and the rover. I moved it from DroneKit onto MAVSDK, kept the v1 Runner/StateMachine surface, and added a v2 async API. Current release is 1.4.11.

• Testbed MAVLink filter only allows takeoff, land, goto, and yaw (20–100 m takeoff, land within 5 m of home).
• SET_MODE is blocked; GUIDED has to be held with MAV_CMD_DO_SET_MODE. LOITER drops the E-VM link.
• Safety checker belongs on the C-VM, not localhost or the OEO console.
• Timed set_velocity was killing newer commands; bearings ignored latitude (~51° vs 45° at 35.7°N).
• Multi-vehicle ZMQ dropped transitions until JSON + ACK + wait_for_peers.

• The filter defines the API as much as the SDK does.
• SITL is necessary and not sufficient.
• A compat layer plus a second API beats a flag-day rewrite.

I wrote PyAerial as the successor to airstrik.py. It decodes ADS-B with pyModeS, evaluates Shapely geofences, and splits tracking (pyaerial run) from a read-only portal (pyaerial web). airstrik used circular filters, MongoDB, and Kafka; this stack is Redis + SQLite + webhooks.

• CPR decode needs a home reference that is not the geofence, plus 0.5° jump rejection.
• Multi-receiver merge on RSSI and Hamming distance; heading ranges wrap.
• Curved ETA can miss a fence when the turn circle blows up — take min(curved, straight).
• Dwell/hysteresis so brief constraint matches do not spam.
• Portal is read-only; the tracker claims Redis with a heartbeat.

• ADS-B is a wire-format problem before it is a map problem.
• Live vs history wants two stores, not one database doing both.
• Geofence rules should be explicit (alt, speed, ETA) — not implicit 'inside the polygon.'

Dockerized field service: Kafka position/velocity in, VAPIX pan/tilt/zoom out, recordings to the camera SD card as Matroska. Stock sensecam-control had PTZ but not record, so I forked sensecam-control-record for start/stop/export. YAML holds site pose, FOV scale, and drone size; dry-run works without a camera.

• Pan is a spherical bearing; tilt is elevation; zoom maps a Q-8615E FOV model into 0–9999 steps.
• Camera north-offset has to wrap into AXIS ±180° pan.
• Optional velocity lead: NEU vs a possible NED convention (sign-flip on vz).
• Minimum pan/zoom steps so the PTZ is not spammed; stale telemetry parks the camera.

Closing a telemetry bus onto a physical PTZ is as much field ops as math: Digest-auth CGI, host-network Docker, and a library that can actually start a recording.

Python CLI for equity research: FactSet FQL snapshots or timeseries, YAML/Pydantic configs, a sandboxed formula engine, and Rich/Excel export. A planner batches FQL calls, then Python metrics run in topological order. FDT mode scores a name against peers into Bull/Base/Bear multiples.

• Formulas needed to be useful without becoming a code-execution hole (AST denylist + allowlisted builtins).
• Interdependent metrics: DAG, cycle path, typo suggestions — before any FactSet call.
• Missing FQL values become typed N/As instead of poisoning a sheet.
• Holdings vs timeseries needed different config models; some formulas need a second-pass fetch.
• Runs only inside FactSet's Programmatic Environment.

Treat research tooling as a product: validate the config language first, then plan → fetch → derive → export.

Slash lobbies, matches in threads, and a plugin API so the same play() runs live, in replay, and on a CLI. Postgres 16 holds guilds, matches, the move log, and live-match integrity. Games are in-process plugins — fault isolation, not a security sandbox. Builtins: Tic-Tac-Toe, Connect Four, Chess (incl. Chess960), Coup, Mafia, Spyfall, Liar's Dice.

• Live matches resume after a restart by replaying the move log; abandon if the plugin build changed.
• Lobby/thread timeouts and signed msgpack custom_id routing were the long tail.
• Hang caps and a thread pool keep a stuck play() off the event loop.
• Public invite exists; the bot is not hosted yet.

A game host is a durability problem: deterministic simulation, an append-only log, and 'live' as a first-class match state.

Content lives in site.toml, templates render it, GitHub Pages serves it. The background is its own engine — twelve modes with pointer interaction and a live parameter panel.

• Keep the sim smooth behind a frosted column.
• Honor reduced motion; persist mode and params.
• Do not steal pointer events from the page.

A portfolio is more convincing when the page itself is something you can operate.