Shaaf Kashan

Skills

Five themes, each with a point of view and the skills that back it. Every skill names the case study, module or credential that shows it.

AI architecture

Decide what the model is allowed to see and do before deciding what it can do.

Local and cloud model routing

In FRIDAY's first version, a routing layer sent each request to a local Llama 3.1 8B model or to Claude. Context was tagged private by hand, private requests were forced local, and the cloud path refused by default.

Evidence: FRIDAY; Claude Certified Architect (Professional)

Retrieval over personal notes

Embeddings with bge-small and a LanceDB vector store grounded FRIDAY's answers in my own notes. The same set of notes drove both the answer and the privacy decision, so the two could not disagree.

Evidence: FRIDAY

Speech to text

Voice input and a wake word were handled by faster-whisper in FRIDAY's first version, running on the Mac rather than in a cloud service.

Evidence: FRIDAY

Machine learning

Benchmarked linear regression, random forest, XGBoost and a neural network for property valuation. XGBoost won at RMSE 0.609 and went into the product behind a Flask service.

Evidence: TerraGrid (A+); CS3002 Artificial Intelligence

Deploying into real environments

A system earns access to real data by keeping its boundaries, encryption and audit trail in place from the first commit, not by adding them last.

Guardrails built first

The privacy and routing boundary went in before any feature, and cost was capped by a flat subscription rather than a metered API. I also learned to check that a tested rule is wired into the running app.

Evidence: FRIDAY

On-device models

A language model that runs on the phone, so a user's reflections never leave the device. Built solo in React Native, not released yet. FRIDAY's first version ran Llama 3.1 8B locally on the Mac.

Evidence: Mihwar (on the phone); FRIDAY (a local model on the Mac)

Encryption at rest

FRIDAY's first version kept conversation history in a SQLCipher-encrypted store.

Evidence: FRIDAY; CS3609 Cyber Security

Tamper-evident audit logs

FRIDAY's first version kept an HMAC-chained, tamper-evident record of actions the assistant took.

Evidence: FRIDAY

Security headers and CSP

This site ships a content security policy and a set of security headers, with a test that keeps the two server paths in step.

Evidence: This site (api/_securityHeaders.js, header-parity test)

Production engineering

Put guards in front of the release, because the failures that hurt are the ones tests and a build cannot see.

Deploy guards

A pre-push gate runs lint, tests, build and a host simulation that boots the built app the way Hostinger does. It exists because two outages were host behaviour that tests and the build could not see.

Evidence: This site (verify:host, .githooks)

Express and SQLite backend

An Express server and a better-sqlite3 store in WAL mode behind a small Redis-style interface, for the guestbook, admin inbox and durable rate limits.

Evidence: This site (api/_kv.js)

Rate limiting and spam control

The contact endpoint has a hidden honeypot field, a per-IP rate limit, a per-recipient auto-reply cap and a daily send cap, with a mailto fallback when the backend is down.

Evidence: This site (api/contact.js)

Testing

A Vitest suite of 600+ tests, Playwright end-to-end checks with axe accessibility scans, and gzipped bundle budgets with Lighthouse CI.

Evidence: This site; CS3003 Software Engineering

Desktop app packaging

FRIDAY's first version was a Tauri and React desktop app talking to a Python daemon over an authenticated local connection.

Evidence: FRIDAY

Strategy and risk

Decide what to build, and how it will be governed, before deciding how to ship it.

Solution architecture

Designing a system from requirements to components, integrations and trade-offs, and writing the decisions down.

Evidence: FRIDAY, TerraGrid; CCAR-P; CS3003

Integration and service boundaries

Deciding where systems connect: FRIDAY's routing layer, and TerraGrid's chain gateway, which isolates blockchain complexity from the rest of the app.

Evidence: FRIDAY, TerraGrid

Problem structuring

Turning an ambiguous problem into options and trade-offs that a non-technical decision-maker can act on.

Evidence: Case studies and decision records on this site

Technology strategy and foresight

Roadmapping, scenario planning and adoption analysis, plus a structured review of an emerging technology (orbital cloud data centres) covering maturity, prospects and risks.

Evidence: UCL SPCE0032, SPCE0034

Risk, reliability and resilience

Failure modes and resilience in large systems, applied in an assessment of the NHS Federated Data Platform.

Evidence: UCL SPCE0025

Systems thinking

Modelling how the parts of a complex system depend on each other, from the systems engineering and engineering management side of the MSc.

Evidence: UCL SPCE0023

Delivery

Technical work only lands when someone who is not technical can act on it.

Agile and Scrum

Ran sprints and a backlog with rotating Scrum Master and Product Owner roles in a group build.

Evidence: WealthSphere (A*), CS2001; PSM I

Project management

Planning, scheduling, risk and lifecycle management for technical projects.

Evidence: UCL SPCE0003; CS2002

Explaining technical work

Supported workshops for 150+ students and turned complex ideas into practical steps, which is the same skill as presenting a decision to a non-technical audience.

Evidence: Brunel Associate Teaching Assistant role

Languages

Python (FRIDAY, TerraGrid), JavaScript and React (this site, FRIDAY, WealthSphere), Java (SWIFTBOT, TerraGrid, WealthSphere), Spring Boot (TerraGrid, WealthSphere), SQL (MySQL, SQLite, SQLCipher) and Solidity (TerraGrid, on Ethereum's Sepolia testnet). Also Git, GitHub and Postman.

Learning next

Enterprise architecture. I have studied the whole-estate view of capabilities, standards and governance through the MSc, and I have not practised it in a job.