Operations leader with 28 years of experience across high-volume customer support,
workforce scheduling, and service delivery with ByteDance, TTEC, Uber, Vodafone and
airberlin (as a client of Aero Service Egypt). Combines deep operating-floor judgment
with hands-on skill building software tools that solve daily operational friction. Proven
track record of streamlining team workflows, supervising operational teams, and
introducing practical, human-guided automation to contact centre operations.

Skills

• Operations & Workforce Leadership: Workforce Management (Erlang C, SLA tracking, schedule adherence, forecasting), Contact Centre Queue Leadership, QA/AQA Frameworks, Incident Escalation, Zendesk, Salesforce Service Cloud.
• Team Supervision: staff reporting to me across four roles, up to 15 at peak; daily operational supervision; coaching and floor support; objective setting; occasional written reviews; performance reporting to senior management.
• Applied Automation & Tooling (evidenced by self-built projects — see Selected work samples): Workflow Optimization, SOP Structuring, Human-in-the-Loop Controls, Process Automation, Python, FastAPI, Desktop Application Design, System Integration Troubleshooting.

• Helix Prime, the flagship: CI green on main — run 37180853667 at a2446b2, 2026-10-04. The preceding code run 36917043028 (head a710804, 2026-10-01) reported 1,951 tests passed, 0 failures, 19 deselected at 86.91% coverage; all six engines drive a real computation end-to-end. Snapshot 2026-10-04.
• The portfolio: one person's self-funded work. Where it is unfinished, the documents say so.
• Looking for one design partner for a shadow-mode pilot. The pilot protocol defines the boundaries; the evidence pack shows what exists today.

A narrated 95-second walkthrough of the portfolio — the core, the components, and the evidence standard. Watch the film Narration is a synthetic voice.

I spent twenty-one years on the operational floor, and the lesson that survived
every one of those roles is that the hard part is never the model. It is the handover:
who owns the decision, what evidence supports it, and what happens when the system is
wrong.

So everything here is built on one rule: on missing input or missing authority, the system refuses rather than guesses. It is audited so any action can be replayed, and it is fail-closed by default. A
generative model may help a person think, but it never sits alone in the path that
executes.

Twenty-eight years in operations. Operations leader with 28 years of experience
across high-volume customer support, workforce scheduling, and service delivery with
ByteDance, TTEC, Uber, Vodafone and airberlin (a client of Aero Service Egypt) —
combining operating-floor judgment with hands-on skill building software tools that
solve daily operational friction.

Twenty-five of those years — 1998 to October 2023 — ran across contact-centre operations, aviation ground operations, and international trade, progressing from frontline and technical support into workforce management, process design, and team supervision.

Staff have reported to me in four roles — in three of them individually, and in the fourth to the function I was part of. Up to fifteen people at peak. I have directed daily work, monitored quality, coached, set objectives, run occasional written reviews, and produced performance reporting for senior management. These were operational reporting lines: I was not their HR line manager, and I do not describe it as managing a team.

Independent Developer — 2026 – present. Building governed, evaluated AI systems independently. Six public repositories, including Helix Codex / Helix Prime and the LIVE Support Assistant, plus a published set of work samples. No client, no revenue, no team and no funding — details and evidence in The portfolio below.

TikTok LIVE (ByteDance), Cairo, Egypt | Sep 2025 – Mar 2026 | Multilingual Customer Support & Operations Lead (TL2).
At TikTok LIVE I noticed agents losing time searching fragmented, fast-changing SOPs,
built a lookup assistant, and ran a one-month live pilot with one colleague. Quality and
handle time held steady throughout; the pilot was not instrumented, so no quantitative
impact was established. I operated within the TL2 function, to which TL1 agents
reported, with on-call escalations routed through the team.

Career transition. I accepted an offer from Etihad Airways for a Senior Premium
Guest Relations role in Cairo and resigned to transition. When Etihad subsequently
cancelled the Cairo office launch, I pivoted to full-time independent software
development. The first tools in this portfolio are dated May–June 2026 in their
repositories.

Four small tools — Erlang C forecasting, real-time adherence monitoring, churn-risk
scoring, and an intake-to-readiness pipeline — converged into one idea: Helix
Codex, an accountable AI operating organization. Not an autonomous agent. An
organization with a constitution, named roles with bounded authority, evidence
trails, and a human at every consequential boundary. Helix Prime is its operations
core.

This work is maintained by one person, with no team and no funding. It has not been
externally audited and it has not made revenue. Where it is unfinished, the
documents say so.

Helix Prime is the operations core of Helix Codex. It is a local-first platform that
owns the parts that must be governed in one place:

• Identity and authentication, with role-based access control (RBAC) that defaults to deny.
• The Erlang C forecasting core — the same staffing maths the May–June 2026 building attempts first sketched.
• The CRM, the workflow engine, and the event-sourced memory that records what happened and why.
• The fail-closed governance gate that evaluates every submission before execution: cost against the acting role's approval limit, data classification against what the role may read, confidence, and engine ownership. It returns dead-letter, awaiting-approval, or executing — and on a miss it refuses, it does not guess.

Everything else in the portfolio is a satellite of this core. Some are components that supply a capability Helix Codex would use. One is a commercial vertical built on the framework. Four are the building attempts whose thinking was absorbed into Prime itself.

Two distinct relationships are in play, and the difference matters.

Code relationship. Every repository in this portfolio is a separate codebase by design. Helix Prime enforces "versioned sibling-service event contracts with no cross-repository imports" — the services talk through published contracts, not shared source. That separation is an architecture decision, not a gap: a tangled monolith would have made the fail-closed boundary harder to prove, not easier. Where a repository depends on the framework, it does so as an external client over those contracts, never by embedding Prime's code.

Data relationship. Most of these are not wired to each other yet. Where a data flow is real and shipping, it is named as such in the details below. Where it is a design intent — the shape of a future pipeline — it is labelled "designed to; not yet wired" so a reader does not mistake a plan for a live feed.

Helix Prime — the operations core. Owns identity and RBAC, the Erlang C forecasting core (pinned by 42 reference tests), the CRM, the workflow engine, the event-sourced memory, and the fail-closed governance gate; nine defined role seats, and dispatch.py is a stub. Pre-pilot. CI green on main — run 37180853667 at a2446b2, 2026-10-04; the preceding code run 36917043028 (head a710804) reported 1,951 tests passed, 0 failures, 19 deselected at 86.91% coverage; all six engines drive a real computation end-to-end; 9 production-only gates red by design; release v1.1.0. Snapshot 2026-10-04.

Helix Education — component: learning engine. Independent repo. Designed to supply the learning foundation Helix Codex would use; not yet wired into the core. Alpha; 465 tests passed on CI (run 36807198469, head 4bd6c02, 2026-10-01); MIT; release v1.1.0. Snapshot 2026-10-01.

Study Studio — component: local-first tutor. Independent repo. No pipeline wired to Prime today. Working; 47 suites / 986 tests; 85.66% statements / 75.13% branches; CI green on main (run 36970701975, head b2799a3, 2026-10-02). Snapshot 2026-10-02.

L&D Command Center — component: desktop workstation. Independent repo. No data flow wired to Prime today. V1 ship in progress; the E2E smoke run was mixed (window launch blocked by a tkinter environment problem; generation exceeded the 60 s budget on that hardware; two integrations untested). Snapshot 2026-09-23.

LIVE Support Assistant — component: explainable support. Independent repo, no shared code or runtime with Prime. Its operational signals could inform Helix Education's curriculum — designed to consume that feed; not yet wired. Prototype; CI green on main (run 36979169955, head 0b40b06, 2026-10-02); designed for a shadow-mode pilot; six prerequisites listed, none met yet; on-device gate 10/10 checks; a contradictory corpus is refused at publish with HTTP 422; backend suites pass but are not wired to the standalone client. Honest limit: at 48 procedures, in-scope recall is 27.8% (84/302) at margin 0.18; 6 of 52 out-of-scope public questions answered at margin 0.18; the deployed 39-ticket parity run found 6 unsafe at margins 0.1869–0.4175. Snapshot 2026-10-02.

Blue Waves — vertical: content studio. Built on the Helix Codex framework; consumes Prime as an external client over its contracts, never embedded. Live client relationship to Prime (HTTP/contracts). Pre-revenue; CI green on main (run 36970381279, head c2dfc0c, 2026-10-02); 167 tests passed (151 project + 16 vendored, re-measured 2026-10-02); the cockpit E2E flake was addressed (client HTTP timeout 30 s → 180 s). Snapshot 2026-10-02.

The four 2026 building attempts — WFM Forecasting Calculator, RTA Command Center, CX Sentiment Sentinel, Dynamic Ops Automation Engine. Historical project context, not audited evidence; two are not runnable as committed. Their Erlang C maths, adherence thinking, KPI-decay risk scoring, and tenant/fail-closed configuration became Prime's WFM, RTA, CX, and B2B engines respectively — conceptual lineage, no shared code. Snapshot 2026-06 / 2026-09-27.

Three self-contained exercises, published in work-samples. Synthetic data only; each runs locally with no cloud service and no paid API. Each one states its own limitations, including the results that reflect badly on it.

• SQL — a synthetic contact-centre dataset on SQLite: two-table joins with measured cardinality, aggregation with HAVING, multi-step CTEs, window functions including a rolling frame, and an index whose effect on the query plan is measured rather than assumed — a full table scan becoming a covering-index search. Eight checks, rerunnable from a clean setup with no dependencies.
• Evaluated RAG — a synthetic policy corpus with a labelled 36-case evaluation set across answerable, paraphrased, conflicting, outdated and out-of-scope queries. A retrieval-only baseline produced 16 unsupported answers; a deterministic decision gate removed all 16, at the cost of refusing questions it could not resolve. 105 tests. The published failure analysis includes where retrieval does not work — paraphrased questions score 2 of 8.
• Container and CI (inside the RAG sample) — Docker and Docker Compose with health checks, structured JSON logs, correlation IDs, query-text redaction by default, request timeouts and a rollback runbook. Six CI steps — format, lint, type checking, tests, evaluation report and image build — all passing. No cloud deployment was performed.
• Consulting proposal and estimate — an illustrative proposal containing a discovery agenda, a no-AI baseline compared against retrieval and generative options, scope and acceptance criteria, a bottom-up man-day estimate with a full assumption register, and conservative/base/upside ROI scenarios expressed as formulas. Explicitly labelled as not a real proposal.

• Local-first development. Cloud only when a customer's value justifies it.
• Governance before autonomy.
• Evidence before claims. Verification requires command output that survives review, not a summary.
• A human at every consequential boundary.
• Memory with provenance, classification, and tenant isolation.
• Improvement through evaluated proposals, never silent self-modification.
• Small, measurable pilots before large infrastructure.

One design partner for a shadow-mode pilot. If you run a contact
centre, an academy, or any operation where an AI-assisted decision has
consequences, I will deploy a governed pilot in shadow mode: the system proposes,
your people approve, every action is replayable. The pilot
protocol
defines the boundaries; the evidence
pack
shows what exists today.

Roles. Contact centre operations, workforce management, and AI-implementation
roles — remote or freelance. I value teams that treat honest "not-ready" signals,
auditability, and local-first privacy as features rather than blockers.

Self-directed study. No certification is claimed for any item below.

• Cloud and LLMOps fundamentals — deploying and operating models, not only building them.
• Workflow automation — n8n and UiPath.

Hatem Ismail Shalaby — Contact Centre Operations & AI Implementation Lead | WFM & CX Transformation

• GitHub: HatemIsmailShalaby1979
• LinkedIn: hatem-shalaby-202902127
• Email: hatemshalaby2025@gmail.com
• Education: BSc Managerial Sciences (Computer Section), Sadat Academy for Management Sciences (2005); Business Analytics Nanodegree, Udacity (2021)

Based in Al Obour City, Al-Qalyubia Governorate, Egypt.

MIT