A grayscale coastline at dusk: fog-layered hills behind waves breaking over dark rocks.

Allston Fojas

Systems Software Engineer, Core OS Power and Performance at Apple

About

San Francisco Bay Area · B.S. Data Science, University of California San Diego

I am a systems software engineer at Apple on the Core OS Power and Performance team. I led the end-to-end design and implementation of on-device telemetry for a next-generation WiFi and Bluetooth chipset, which made it possible to compare battery performance against prior chipsets quantitatively across large device fleets. The telemetry shipped with the chipset, and it turned battery questions that had been argued from intuition into questions answered with fleet data. The implementation is Objective-C and C++ across iOS, macOS, and watchOS. The analysis side is Python, Jupyter, and Spark.

Getting that built meant driving alignment across the WiFi, Bluetooth, Privacy, logging infrastructure, and analytics teams.

Outside of work I build decision systems, and the architecture is the same in every domain. A versioned scoring rubric is the single mutable surface, read at call time. Every heuristic change lands in an append-only log. A two-layer model compares what the system predicted against what actually happened, re-scoring is a deterministic tested function rather than a judgment call, and every rubric mutation stops at a hard human-approval gate. I have run it against a housing search and a content production pipeline.

I run a local inference lane on 24 GB of Apple silicon so that agentic coding keeps working with no network and no third party holding the data. It serves open-weight Qwen and Gemma models over OpenAI-compatible and Anthropic-compatible routes, so the same harnesses I use online, Claude Code and pi, drive it unchanged offline. Getting tool calling to work there meant root-causing a failure to a model dropping its sentinel tags, proving it a model fault rather than a harness fault by A/B against a second model, and shipping a parity gate so it cannot regress.

The housing system is the one with an outcome attached. It scored 531 candidate listings across 26 discovery runs against a deterministic, versioned admission gate, ranked them with a dual-anchor commute engine built on open routing data, and shipped a public board. I signed a lease on a property the system surfaced and vetted.

Skills

Programming Languages

Python
Objective-C
C++
SQL
Swift
JavaScript
HTML
CSS
Java

Development Tools

Git
Jupyter
Spark
Streamlit
Tableau
Jenkins
Kubernetes
Convex
Vercel
Bash

AI Agent Tools

Claude Code
Codex
LangGraph
Claude API
OpenAI API
Gemini
MCP
OpenCode
Pi

Experience

2022-07 - Present · Cupertino, CACurrent

Systems Software Engineer, Core OS Power and Performance

Apple

Led end-to-end design and implementation of telemetry for a next-generation WiFi and Bluetooth chipset, enabling quantitative battery performance comparisons against prior chipsets across large device fleets.
Drove cross-functional alignment with the WiFi, Bluetooth, Privacy, logging infrastructure, and analytics teams.
Implemented on-device telemetry pipelines in Objective-C and C++ for iOS, macOS, and watchOS.
Designed and executed large-scale data analysis workflows using Python, Jupyter, and Spark.
Built agentic AI workflows to make investigations of key metrics more efficient.

2020-08 - 2022-06 · San Jose, CA

Software Engineer

Cisco

Developed performance test scripts in Python and ran them against the team's Kubernetes cluster.
Implemented unit tests for the team's services in Go, with CI/CD through Jenkins.
Built data visualization dashboards in Grafana and analyzed test logs in Kibana.

2020-06 - 2020-08 · San Jose, CA

Data Engineer Intern

VIZIO

Created a location weather API driving product recommendations surfaced on VIZIO TVs.
Built the API and recommender with Python, Pandas, scikit-learn, and Scrapy.

2019-06 - 2019-08 · New York, NY

Software Engineer Intern

American Express

Replaced a $1M vendor relationship with an Excel file normalizer that standardized incoming formatting.
Worked with Product Managers and the Business team to meet the business use case.

Projects

Waypoint - agentic trip planner

Apr 2026 - Jun 2026

An orchestrator that dispatches four parallel Claude research agents to produce a costed, sequenced trip itinerary.

248 tests. Built across a five-hackathon cascade and presented at the Andreessen Horowitz office in San Francisco.

Four parallel Sonnet research agents - points of interest, hotels and flights and restaurants, photography, transit - each a multi-turn web-search loop terminating in a structured submit tool call.
A provider-abstraction layer that allows swapping the generative-media model without touching the pipeline.
A prompt-injection input sanitizer, per-session cost telemetry, and graceful partial failure when one agent returns nothing.
A second documented research path builds a prompt for Perplexity Computer and parses its output back - copy-paste through the UI, not an API integration.
Presented at the a16z, Overshoot, Fal and Mux Video hackathon, May 2026.

Python 3.12
Anthropic SDK
FastAPI
Jinja2
fal
Mux
FFmpeg
SQLite
pytest
Playwright

Product page

Automated housing-search decision system

2026

Scored 531 candidate listings across 26 discovery runs against a deterministic, versioned admission gate. Signed a lease on a property from the board.

Signed a lease on a property the system surfaced.

Deterministic, versioned admission gate: hard constraints reject before any scoring runs, so a rejection is always explainable.
Dual-anchor commute engine on Nominatim and OSRM, with no paid routing API.
Shipped two surfaces: a static GitHub Pages board and a Next.js, TypeScript, and Convex application.
131 tests across the Python pipeline and the TypeScript app.

Python
TypeScript
Next.js
Convex
Leaflet
OSRM
Nominatim

Live board Source

Execute-and-learn decision loop

2026

A reusable execute-then-learn loop applied across two unrelated domains: housing search and content production.

The scoring rubric is the single mutable surface and is read at call time, so behaviour changes are data changes rather than code changes.
An append-only heuristic change log makes every past scoring decision reconstructible.
A two-layer prediction-versus-outcome model separates what was predicted from what was observed.
Re-scoring is a deterministic tested function; every rubric mutation requires human approval.
Four consecutive retrospectives declined to change a weight because the outcome sample was too thin to learn from.

Python
Markdown-as-data
LanceDB

Also

Concurrency-safe git layer for multi-agent workspaces

2026

704 lines of Python that let many agents commit to one repository without clobbering each other.

Advisory flock with explicit handling for deleted-lockfile and split-inode races.
An index-health gate that blocks commits during a filesystem-eviction state which would otherwise commit mass phantom deletions.
Manifest-scoped staging, so parallel agents cannot stage each other's in-flight work.
Bounded rebase-retry that aborts loudly instead of force-pushing.

Python
Git

Headless content-production pipeline

2026

Three Python packages, roughly 2,900 lines and 211 tests, that turn a semantic search over raw footage into a rendered cut.

Search-driven edit-decision lists reached 9 of 9 top-1 retrieval accuracy on the evaluation set.

Semantic footage retrieval over TwelveLabs Marengo 3.0 and Pegasus 1.2, across both visual and audio signal.
An FFmpeg render engine driven by a typed beat contract rather than imperative edit steps.
A Remotion transparent ProRes 4444 alpha overlay track composited over the base render.

Python
FFmpeg
Remotion
React
TypeScript
TwelveLabs

Local-first inference lane with a fail-closed privacy guard

2026

An on-device model lane on 24 GB Apple silicon, with a three-layer guard that fails closed so private corpora do not leave the machine.

Ollama and MLX serving both OpenAI-compatible and Anthropic-compatible routes behind one interface.
A three-layer allowlist guard, not a denylist, that fails closed before reading any user-supplied parameter, which makes it injection-resistant.
The huggingface route is deliberately excluded from the allowlist because the router silently forwards it to cloud.
Root-caused a tool-calling failure to a model dropping sentinel tags, proven a model fault rather than a harness fault by A/B against a second model, then shipped a parity regression gate.

Python
Ollama
MLX
Apple silicon
Qwen3.5
Gemma 3
Claude Code
pi

Agent skills library (private)

2026

A private library of composable agent skills: the operating manual for the systems above.

Each skill encodes one workflow as an explicit procedure with its own verification steps, rather than relying on a prompt.
Skills compose: planning, test-first implementation, and independent review are separate skills chained by a controller.

Markdown
Python
Bash

Research vault showcase

April 2026

A public walkthrough of the retrieval architecture behind a personal research corpus.

Hybrid retrieval: vector search with HNSW cosine, BM25 lexical search, and reciprocal-rank fusion over the merged result set.

Python
LanceDB
Tantivy

Live

Earlier work, 2019 to 2020

Predicting Disease Risk

Apr 2020 - Jun 2020

Supervised prediction of individual risk for three diseases: coronary artery disease, Alzheimer's and diabetes.

Accuracies of 0.91, 0.89 and 0.88 across the three disease models.

Predicted risk for three diseases from one body of modelling work: coronary artery disease, Alzheimer's and diabetes.
Built as the UC San Diego Data Science capstone in the genetics domain.
Delivered as reproducible notebooks.

Python
pandas
NumPy
scikit-learn
Matplotlib
Seaborn

Source

Weather Forecast IoT Device Map

Apr 2020 - Jun 2020

A web app mapping IoT weather sensors and testing whether forecast accuracy varies by city.

Determined that forecasts for Sacramento are about 7% more accurate than for San Diego.

A web app spanning an ESP32 sensor device, a MySQL store and a browser front end.
Compared forecast accuracy between two cities and quantified the gap.