Gil Raitses
Embedded delivery
Three programs I built and shipped, and how each one maps to embedding with an engineering team and owning delivery end to end.
I take a research problem from raw data through a working, deployed tool, get it adopted by the people who use it and run the agent work behind it so it stays coherent across many repositories and sessions. The same instincts carry to deploying a product inside a customer's environment. This page links the surfaces you can open directly.
orcast
orcast is an evidence-gated encounter forecasting program for Southern Resident killer whales. It pairs biologging and sightings data with validation gates, a workbench for researchers and a public interactive module already online as Orca Strike while the wider platform finishes.
Where it maps: I build and ship a research product, deploy an interactive surface into a running environment and design the integrations that carry measurement into a team's workflow.
wavves
wavves is the method I run agent work through. It assigns a cast of agent lanes, each a role with explicit inputs, outputs and an escalation rule, records checkpoints as the unit of memory and runs an alignment check that flags when the structured record and the human-readable views diverge. A recent waveset layer lets the orchestrator charter a campaign and dispatch waves of parallel subagents to a fresh orchestrator, so a single run can fan out without losing the audit trail.
Where it maps: it is a working answer to the question a customer asks when they scale agents, namely how to trust and audit what the agents did. It doubles as a measurement surface for agent behavior across real, multi-repository work.
aimez.ai
aimez.ai is an applied research program that takes public traffic-camera frames through computer-vision models into a crowd pressure field, then into pedestrian routing that respects measured burden. It carries the work end to end, from raw field data to a public site with notes, figures and validation briefs.
Where it maps: it is field deployment of the messy kind, turning ambiguous real-world data into a working system, with customer-facing documentation that explains the process to a non-specialist.
The throughline to the team's objectives
Stated plainly, here is what forward deployment needs and where I have already done it.
- Deploy and integrate in a real environment. Orca Strike is live as the public interactive module for orcast, with the wider research workbench still under active build.
- Earn adoption against real resistance. I built retrovibez, a larval reverse-detection pipeline, and a mechanosensation group adopted it after I visualized every detected event so they could check the calls, demoed it in a lab meeting, trained new researchers and wrote the install and usage docs.
- Make agent work reliable at scale. wavves keeps long multi-repository agent work coherent and auditable, which is the surface a team needs to trust agents in production.
- Translate an ambiguous problem into a shipped system. aimez.ai runs from raw camera data to a routing tool and a public, documented site.