Where I’ve built, and for whom.

Promoted from AI Engineer Intern after four months.

• Architected a nine-agent course production pipeline, producer, research analyst, curriculum architect, deck writer, assessment designer, lab engineer, QA reviewer, TA agent, video producer, turning a course topic into an instructor-approved package. Every agent has a written brief, a model tier and folder-scoped write permissions.
• Enforced a forward-only state machine with three human approval gates and a 940-line append-only approval log that records a real Gate 2 rejection, not only sign-offs. An 18-check QA reviewer blocks the final gate and is permitted to find defects, never to fix them.
• Wrote the founding architecture of a query-to-dataset factory, deterministic streaming pipeline, 11 declared assertions, a seven-lens adversarial QA panel in which each finding must survive two of three independent refutation attempts, and a compliance gate that can only permit or block, then ran the first end-to-end build by hand, producing a 22,195-record dataset identical to my hand-built reference. The system streams ~9.8M registry rows per national build and has delivered datasets of 409,848, 73,394 and 22,195 records.
• Wrote the per-agent security model both pipelines run under: folder-scoped read/write permissions, sandboxed code execution for the lab agent, an untrusted-input policy for the student-facing TA agent, and, on the provider factory, the decision to build no send path at all, deliberately and by build order.
• Routed model cost by judgment density, heavy reasoning models only for judgment-dense agents; the high-volume TA agent on the cheapest capable model, with accuracy carried by a RAG grounding pack rather than model size.
• Shipped the initial codebase, backend wiring and security work on a production HRIS for South Asian SMEs: Fernet field-level encryption on national identity numbers, TOTP MFA with hashed recovery codes, security headers and asset caching, the Entra SSO design, and an audit row written inside the same transaction as every mutation. The system runs to 12 domain packages, 123 REST endpoints and 215 defined backend tests.
• Contributed to an LLM agent pipeline that generates, evaluates and archives quantitative alpha factors from 1-minute equity bar data, with self-validating submission notebooks carrying hard data-quality gates and full AI-provenance audit trails. The top AI-generated factor scored 0.826 on the platform leaderboard, team best, as of 2026-07-27, before the 2026-08-02 freeze.

• Python 3.12/3.13
• Claude Code subagents
• Claude Opus 5 / Sonnet 5 / Haiku 4.5
• Google Gemini
• RAG grounding packs
• FastAPI

• Architected a modular crypto futures trading system as a seven-layer pipeline, data ingestion, feature engineering, market-state detection, multi-strategy signals, signal fusion, risk management, execution, with separation of concerns strict enough to hold backtest/live parity.
• Delivered two further production AI systems, Market Pulse and a Reconciliation Agent, on event-driven architecture with structured LLM output and database-level data-integrity constraints.
• Researched token usage, token economics and token-optimization strategies across the leading LLMs to guide cost-efficient model selection and routing.

• Python
• Event-driven architecture
• JSON-schema LLM output
• Backtest–live parity
• LLM token accounting

Started 1 September 2026. Runs to 31 December 2026, alongside the Symufolk role. Outcomes will be added as the programme produces them.

• Migrated 300 GB of reporting from Qlik Sense to Power BI, rebuilding the data models and the ETL/cleaning logic rather than porting them.
• Delivered eight publicly reviewed engagements, dashboard development, data analysis and data visualization, for clients in five countries (Pakistan, the UK, Canada, the UAE and the US), one of whom came back for a second order. Order values $50–$200, delivery 1–6 days.
• 4.9/5 across those eight reviews (seven 5-star, one 4-star), at Level 1 Seller, a platform-awarded tier, not a self-reported one.

• Power BI
• Qlik Sense
• SQL
• Excel
• Dimensional modeling
• ETL / data cleaning

• Led technical workshops on Python, SQL and AI/ML fundamentals for 100+ students.
• Arranged industry tours connecting students to working engineering teams.

• Python
• SQL
• AI/ML fundamentals