What I do
I build the gate that decides whether an agent's work counts.
Regression suites, deterministic replay, escalation rules, and audit trails. You get a command that decides, and a record of what it decided.
# colony-kernel — the kernel I build engagements around $ npm run verify gate typecheck: OK gate lint: OK Tests 130 passed (130) gate claims: OK (130 tests, 4 surfaces agree) gate demo: OK (byte-identical reruns, 5613 bytes) verify: ALL GATES GREEN
Failure modes
| What goes wrong | What I put in |
|---|---|
| The agent reports success and nobody can check it | A test that fails when the behaviour regresses |
| State lives in context instead of on disk | Typed handoffs and checkpoints that a session cannot quietly corrupt |
| Failures are silent, so the same bug ships twice | A failure-mode map and gates that fail loudly on bad configuration |
| Nobody knows when to let the agent act alone | An escalation rule that names the cases a test cannot decide |
| Production incidents leave no trail | Content-addressed run logs an auditor can re-verify |
Engagements
| Engagement | Fee | You get |
|---|---|---|
| Reliability audit | $150–$1.5kfixed fee, a few days | Pipeline review, failure-mode map, prioritized hardening plan, recorded walkthrough |
| Eval-harness buildmost value | $1.5k–$6kfixed fee | Regression suite, machine-checkable gates, escalation rules, audit trail, handover docs |
| Ongoing advisory | from $150/moretainer | Architecture reviews, reliability guardrails, cost routing, priority async support |
| Grounding and cost routing | Scoped per system | Retrieval grounded in approved knowledge, cheapest-model routing with automatic failover |
The first audit is deliberately cheap. If the honest answer is
"you need npm test and nothing else," that is what the call says.
Open source work
| Project | What it proves | Check it |
|---|---|---|
| agent-handoff Python, zero dependencies | Typed handoffs so a long agent session survives a restart. Installs in three commands; ships a normative spec and a conformance suite any implementation can meet. | agent-handoff doctor --strict |
| colony-kernel TypeScript, 130 offline tests | Deterministic mission control: 22-stage state machine, content-addressed approvals, byte-identical replay. Run the mission in your browser. | npm run verify |
| crew-research-council Python | Multi-agent research governance: blackboard state, veto pipeline, role-scoped write rights, so one confident answer cannot pass as a finding. | pytest |
| dspy-lab Python | An offline LM lab: prompt compilation, RAG, graph state, typed tool calls. No API key needed to iterate. | pytest |
| dsh-fallback-chain JavaScript | Free-tier routing: automatic provider failover, quota-aware and cost-aware model selection. | node test |
The rest of the repositories, including the ones behind this page: github.com/faresrafat3.
How I work
- Records over claimsA claim is real only when a verifier can quote it from disk.
- Run it before saying it worksThe command, the file, the port. Never reasoning alone.
- Fail loudMisconfiguration surfaces immediately. No silent skips.
- Deterministic by defaultSeeded runs, replayable state, byte-identical output.
Contact
Tell me what your agent claims, and what would count as proof.
Twenty minutes is enough to tell you whether I can help, and what I would change first.