Projects
Ten things I've built. Some are open-source infrastructure, some are the systems a business actually runs on, some are enterprise platforms deployed at national scale. What they share is that I owned every side of them: the product call, the code, the customer, and the number.
Product Manager
Roadmap, scope, pricing. Deciding what gets built and, harder, what does not.
Forward-deployed engineer
In the code and in the customer's office. I build against what is in front of them, not what the spec says.
Solutions
Turning “we need something like this” into a design, a price, and a plan that still holds once the dates are real.
Business & strategy
Go-to-market, contracts, partnerships, and the number the work has to hit.
Every project below is tagged with the jobs it actually asked for. The ones carrying all four are the ones I'm proudest of, because none of them let me hand the hard part to someone else.
Nodeflux OS
The internal system that puts pre-sales, sales, marketing, engineering and delivery on one shared model of what we sell, what we build, and what is running at our customers.
- — One vocabulary across five departments. Product lines, SKUs, deals, commitments, work items, billing and delivery all resolve to the same objects instead of five spreadsheets that disagree
- — Quotes, pricing, approval chains and payment schedules designed against how the business actually closes and delivers a deal, because I run that process
- — Every architectural decision is written down, including the rejected ones and the three times the model had to be repaired. The expensive mistakes here are modelling mistakes, not coding ones
- — Product owner, architect and engineer on the same system. That is the only way the model and the process stay honest to each other
Khawariz
A quant trading system from market data to live execution, built so a strategy has to earn its result before anyone believes it.
- — A strategy is a config file, not code, so testing an idea costs minutes rather than a release cycle
- — Every parameter combination ever tried is counted across sessions, so the bar a result has to clear reflects the whole search instead of the last command
- — The first strategy through it returned a null, and the null is in the repo rather than buried. The loop is meant to be trustworthy before it is profitable
- — Risk, cost and honesty are the requirements here, and those are business calls before they are engineering ones
Photon
A self-hosted observability platform that goes after Datadog and Grafana on price and footprint. Logs, traces, metrics, APM, RUM and uptime in one binary.
- — Positioned against Datadog, SigNoz and Grafana: one process instead of a stack, flat self-hosted cost instead of per-host and per-seat metering
- — Eight signal types under one roof, so going from a log line to its trace to the service's health is a click rather than an integration project
- — Shipped through to 1.7.0 with a released browser SDK and a host agent
- — The product calls came first. What scale to target, what to deliberately not build, and what the cost story has to be to win a switch
Oasis
The AI agent framework Go deserves — composable, interface-driven, and built for where AI is going.
- — Recursive agent composition — LLMAgent, Network, and Workflow all satisfy one Agent interface
- — No vendor lock-in — raw HTTP to Gemini, OpenAI, Groq, DeepSeek, Ollama, and any OpenAI-compatible API
- — Graph RAG with LLM entity extraction, hybrid retrieval (vector + full-text + re-ranking), and semantic chunking
- — Go-native concurrency — parallel tool dispatch, background agents via Spawn(), and DAG step execution
ix
A sandbox runtime for AI agents — lightweight Firecracker microVMs running a Rust daemon for shell, code, file, and browser ops.
- — Sub-100ms cold VM creation, sub-1ms from pool; shell ops under 3ms from pool
- — Rust daemon runs as PID 1 inside the microVM, exposing HTTP + SSE over vsock
- — Three rootfs tiers — base (~400MB), browser with Chrome (~1.5GB), full with scientific Python (~3GB)
- — DNS-based egress filtering for secure sandboxed execution
nilo
A backend toolkit for Zig built on one idea: your types are the contract, and the compiler is the check.
- — A route is a plain function. The signature produces the URL matching, the parsing, the error responses, and the OpenAPI document, with no second copy to keep in sync
- — Built for people and for coding agents at the same time, which turned out to be one job rather than two
- — 141 error messages held in place by a build step, so a mistake comes back as a sentence while you compile instead of a 500 at runtime
- — Developer experience treated as the product surface. The API is the thing being designed, not a by-product of the implementation
Saki
A native, lightning-fast database client that stays out of your way — built native, not wrapped, and data-first.
- — Tauri v2 + Rust core with a trait-based driver registry — PostgreSQL and SQLite today, more engines on the roadmap
- — Binary IPC for results; a virtual-scrolled grid handles thousands of rows with transactional inline editing and undo
- — CodeMirror 6 SQL editor — schema-aware autocomplete, formatting, multi-statement run, and an EXPLAIN ANALYZE visualizer
- — Capability-aware UI adapts the tree, menus, and tabs per engine; credentials stored encrypted
devrun
A process runner for local development whose logs a coding agent can actually read. Four commands are the whole interface.
- — Starts every service in a repo, keeps each one's output in a plain file, and answers questions about all of them in one command
- — Fixes a workflow problem people had stopped noticing: an agent debugging your app reads one log out of four, then guesses at the rest
- — One static binary at 0.6 MB and 1.2 MB idle memory, with nothing to fetch at runtime
- — The whole design brief was scope discipline. Four commands, no daemon to babysit, no config to learn
Lenz Dashboard
Nodeflux's Visionaire dashboard end to end — a real-time analytics interface and the API backbone that powers it.
- — Streams live video and AI events at national scale across immigration, traffic, and enterprise deployments
- — Unifies recording, analytics, enrollment, and mobile services into a single configurable runtime
- — Real-time operations interface over thousands of concurrent camera and event streams
- — Multi-tenant — per-client configuration packs for distinct government and enterprise programs
AI Analytics Platform
Enterprise computer-vision analytics serving national infrastructure.
- — Real-time video analytics pipeline processing 500+ concurrent camera feeds
- — Alerting and reporting engine for government and enterprise clients
- — Owned the full product lifecycle from concept to national deployment