AI systems & automation / Independent engineering
Selected work / 2026Ishai Chertkoff — AI systems and automation engineering. Ambition, made real.
I’m Ishai Chertkoff.
I build controlled AI workflows, backend systems and cloud automation.
Independent project work. Decisions and evidence open to inspection.
One decision / Lost confirmation
Optional browser illustration. No external actions.
The person behind the systems
I care about what
happens after “done”.
A task can be prepared without being performed. An action can happen without a confirmation. That gap is where much of my engineering work lives.
I build across Linux, cloud infrastructure, backend integrations and automation—turning ambiguous requests into work that can be inspected, continued and questioned.
Request my résuméSelected engineering / 01—03
The decisions behind the work.
Three independent projects. The problem, my contribution and the evidence you can inspect.
RACAI OperatorDurable state & approval boundariesExploreClose
Internal platform · Public engineering notes
State should outlive the conversation.
The problem
Work that starts in chat needs a lasting record of what was requested, approved and performed. The conversation alone cannot be the system of record.
My contribution
I built task and proposal flows, approval and rejection handling, scheduled processes and audit records around a Cloudflare runtime with D1-backed state.
The decision
Keep state outside the conversation. Record a proposal separately from an authorized change, so work can be inspected and continued across interfaces.
WORK REQUEST
↓
STRUCTURED PROPOSAL
↓
OWNER DECISION ──→ REJECT / REVISE
↓
AUTHORIZED STATE CHANGE
↓
AUDIT / EVIDENCE RECORD
↓
FOLLOW-UP OR REVIEWInspect this flow in the public notesTypeScript · Cloudflare Workers · D1 · Web / Telegram
Control-RoomControlled agent executionExploreClose
Historical implementation · Private source
Reasoning is not authority.
The problem
An agent can propose an action without having permission to execute it. Client work also needs scoped records, bounded resource use and protection from duplicate action attempts.
My contribution
I developed bounded planning and tool execution, scoped memory, verification and an operator-controlled executor. External actions became proposals requiring separate approval.
The decision
Before invoking an adapter, claim approved work inside a PostgreSQL transaction. Check approval, enabled adapter type and forbidden categories outside the model’s reasoning.
SELECT ...
FOR UPDATE
“Competing attempts could then observe the claim and back off.”
The public notes describe a transaction that claims an approved proposal and persists an executing state before an adapter is invoked. This is a documented implementation decision, not a disclosure of private source code.
Node.js / Express · PostgreSQL · Python · Agent tooling
ISHAI OS / RAC-stackSource-backed verification & recoveryExploreClose
Active development · Owner-observed bounded workflow
An observation needs a source.
The problem
An agent can mistake a prepared proposal for a completed external action, or an old observation for a current fact. Local memory needs provenance as well as persistence.
My contribution
I developed a bounded, read-only GitHub workflow that checks current pull-request state, validates historical CI artifact digests and stores timestamped, source-attributed results.
The decision
Separate intent, permission, execution and evidence. Preserve earlier observations while recording a newer result, and distinguish a current API read from historical CI evidence.
- Current pull-request read
- Observed: A fresh authenticated API read returned a merged state with an observation timestamp.
- Does not prove: The CI run belongs to the merged head.
- Historical artifact integrity
- Observed: Two ZIP archives matched SHA-256 values recorded in the CI upload logs.
- Does not prove: Provenance or release readiness from a hash match alone.
A public summary of a controlled, owner-observed workflow. The artifacts and runtime records remain private.
Inspect the complete field notePython · SQLite · GitHub API · Local-first operations
Evidence / without the theatre
Inspect the work.
Know its limits.
Different evidence answers different questions. Here is what is available to examine.
- Public implementation
- Python / SQLite task-state example
Inspectable code, separate from the private platforms.
- Locally tested
- 11 Python tests passed · 15 Nov 2026
Concurrency, persistence and invalid transitions. Pinned revision
160e3af. - Public engineering notes
- Three focused case studies
My contributions, consequential decisions and stated limitations.
- Not established here
- Production delivery, uptime or independent platform verification.
Local tests and authored notes cannot establish those claims.
The next conversation
Let’s build
something real.
I’m Ishai Chertkoff. If your team needs someone who can turn an ambiguous problem into a working, reviewable system, let’s talk.
AI systems & automation. Cloud infrastructure. Backend reliability.
My résumé is private. Share the role or opportunity you have in mind, and I’ll review your request before deciding whether to share it.
Tell Ishai what your team is building and where you think he could contribute.