Portfolio · selected work
Each project is shown in the space it actually inhabits — the physical shape of a customer’s workflow, or the mathematical shape of the patterns underneath. Scroll to walk through them.
BeginA $3,000 desktop that runs a full AI assistant on-site — chat, memory, and an audit trail. Nothing leaves the building.
Sized so the chat experience matches what people expect from the cloud, but every byte stays in your office. Pays for itself against a $1,000/month cloud chat bill in roughly three months — after that, the line item is electricity.
An operations platform that ties every cost to a job the moment it enters QuickBooks — so margin is visible mid-build, not at year-end.
Built for any project-based business — construction, architecture rehabs, consulting, field services, agencies, and more. Bills, purchases, payroll, and subcontractor invoices all arrive pre-coded to a job and flow cleanly into the same QuickBooks file your CPA already opens.
An AI assistant that works directly against your CRM and financial data — with a memory that actually remembers.
Memory is treated as a first-class graph: people, accounts, deals, decisions, and past interactions live as connected nodes the assistant can traverse. Every suggestion lands in context, not from scratch.
Search across images and text together. The system finds what customers meant, even when they didn’t have the right words.
Pictures and descriptions are placed near each other in the same feature space. Tuning the system on the customer’s own catalog lifted top-10 results by 20% over the off-the-shelf baseline.
A compact forecasting engine trained on real demand data inside a 10-hour budget.
Time series gets sliced into segments; each segment becomes a pattern fingerprint stored in a fast-lookup library. When a similar shape shows up later, it’s recognized instantly — sized for edge hardware, not just cloud GPUs.
An assistant that scans your machine and installs the private chat experience it can actually run.
Starts from the machine and works backwards. Processor, graphics, memory, storage, and OS all get inspected; a tiered always-on chat + search stack gets installed that’s matched to what the hardware can really run. Same install path on macOS, Windows, and Linux.
A 60-qubit quantum circuit simulated on commodity hardware via mathematical simplification.
The trick is collapsing the structure of the circuit into something small enough for a laptop to compute, then sampling the answer one step at a time. The interesting result isn’t the qubit count — it’s that the right mathematical instinct can replace a supercomputer.
Past work · Accenture Federal Services · 2021–2025
Four years on the hands-on rapid prototyping team inside Accenture Federal Services’ MLE Group. The arc, every time: build something the customer falls in love with, iterate with them, then roll up sleeves and ship it to production — the unglamorous part most demos never reach. Customer names omitted; the outcomes are real.
Co-architected a production platform that combined structured records, semantic retrieval, and an audit-friendly access layer at enterprise scale. The same patterns we now compress and ship to SMBs.
Coordinated the stakeholders who decide whether AI ships at all. Co-authored deployment and use-case guidance for sensitive environments, mapped use cases to clearance levels, and turned a “maybe someday” stance into a working deployment path.
Built the platform plumbing the rapid prototyping team used to ship against customer data — auth, data access, deployment, observability — so an engineer could spin up an experiment in hours instead of days, with a path to production from day one.
Designed and shipped the distributed services behind a multi-million-user record-keeping workflow. Established the observability and on-call patterns that later platform teams inherited as the default.
Designed a search system that joins multiple data sources at ingest time, enriches them with language understanding, and respects fine-grained access controls — so users only see what they’re allowed to, while the index stays unified.
A desktop you can put under a desk that runs a full AI assistant on-site — no cloud bill, nothing leaving the building.
The pattern. Many SMBs want the chat-and-search experience that cloud AI products offer, but can’t send their data to a third party — whether for legal, contractual, or competitive reasons. Existing “run AI locally” advice usually means a laptop running a small model badly. We close the gap.
What we built. A $3,000 desktop sized for high-memory chat, plus a three-tier memory stack: structured records, semantic search, and an audit log of every interaction. The user-facing chat feels indistinguishable from the cloud; the data never leaves the box.
Why it matters. A $3K hardware bill amortizes against a $1,000/month cloud chat bill in about three months. After that the line item is electricity. Because nothing leaves the office, the deployment survives the strictest data policies without an exception.
Scope your data & workflows in one 30-minute call, build & ship the workstation pre-assembled, hand off with a runbook and a recorded walkthrough. Total elapsed time: typically two to three weeks from intro to operational.
An operations platform tied to QuickBooks where every cost lands pre-coded to a job — so margin is visible mid-build, not at year-end.
The pattern. Project-based businesses — construction, architecture rehabs, consulting, field services, agencies, and more — lose visibility the moment a bill arrives in QuickBooks without job context. Bookkeepers spend weeks each month reconciling. Owners only learn a project lost money long after it closed.
What we built. A custom operations platform that sits on top of QuickBooks Online and forces every cost, bill, purchase order, and payment to be tagged to a specific job the moment it enters the system. Estimates from the bidding stage, field costs from the job site, and accounting records all live in one source of truth. The CPA opens the same QuickBooks file they always have — except now every transaction already has job context baked in.
Why it matters. Project managers see margin slipping mid-project and can correct course. Month-end reconciliation work disappears. The example we ship is a $5M construction company that saves roughly $30,000/year in bookkeeping and CPA fees — but the pattern works anywhere work is organized as discrete projects.
Audit your current job-cost workflow in a 30-minute call, map your custom job-coding scheme onto QuickBooks classes/customers/jobs, ship a working prototype in your first month and a production rollout in roughly three.
A full-stack web app where AI assistants work directly against your CRM and financial data — with a memory that remembers.
The pattern. Most assistant demos forget what happened yesterday. That makes them useless for operational work, where every interaction needs to be anchored in the actual state of the business — people, accounts, open deals, past commitments.
What we built. A full-stack web application where an AI assistant works directly against the customer’s CRM and financial systems. Memory is treated as a first-class graph: entities, relationships, and past decisions live as connected nodes the assistant can traverse, not as a flat bucket of text. Each suggestion lands in context.
Why it matters. The assistant survives multi-step workflows that take days or weeks — following up on a stalled deal, chasing a stale PO, prepping a renewal — because it remembers what happened in step one when it gets to step seven.
Map your operational systems (CRM, finance, calendar, comms) in a discovery call, identify the two or three workflows where remembering most matters, ship a working prototype against your real data in four to six weeks.
A search platform that understands both pictures and the words people use to describe them.
The pattern. Customers shop by “the one that looks like this” as much as by exact keywords. Off-the-shelf keyword search misses these queries entirely; the customer leaves without finding the thing.
What we built. A search platform that places pictures and descriptions near each other in the same feature space. A search lights up everything nearby in that space, even when the user’s words and the catalog’s words don’t literally match. The platform is deployed as a managed cloud endpoint so the rest of the system calls it like any other API.
Why it matters. Tuning the model on the customer’s own catalog lifted top-10 results by 20% over the off-the-shelf baseline. The user-visible effect is a search that “just finds what I meant” instead of returning nothing.
Audit your current search experience and miss rate in a 30-minute call, identify the categories where visual + text similarity beats keyword, deliver a tuned production endpoint in roughly six to eight weeks.
A compact forecasting engine trained on real electricity demand inside a 10-hour budget.
The pattern. Forecasting is most useful when it doesn’t live on a leaderboard but on a piece of edge hardware that has to respond quickly under production load. Most published research optimizes for the leaderboard; we optimized for the deployment.
What we built. A compact forecasting engine trained on Australian electricity demand inside a 10-hour compute budget. The model breaks the time series into overlapping segments and learns a fingerprint for each one. The fingerprints are stored in a fast-lookup library — when a similar shape shows up later, it’s recognized instantly.
Why it matters. The interesting question wasn’t “does the model learn?” — it was “is the learned shape useful at serving time?” The benchmark was head-to-head: brute-force lookup vs. the fast-lookup library, on production-sized load.
If you have a forecasting workload that needs to live close to where the data is generated — a meter, a sensor, a POS terminal, a manufacturing line — we can size a model + lookup library for the hardware you actually have.
An assistant that scans your machine and installs the private chat experience it can actually run.
The pattern. Most “run AI locally” advice ignores the machine in front of you and recommends the same setup regardless. The result is either a model that crawls on under-powered hardware, or one that doesn’t use the GPU it could.
What we built. An assistant that starts from the machine. It scans the processor, graphics, memory, storage, and operating system, then installs a tiered always-on chat + search experience matched to what the hardware can actually run. Same install path on macOS, Windows, and Linux. Zero cloud calls in the loop.
Why it matters. The output is a reproducible install, not a Notion doc. Two machines with the same spec get the same setup. A machine with double the memory gets the larger tier automatically.
If you’re trying to roll out private AI across a team with different machines, we can package the installer to match your hardware fleet and your security posture.
A 60-qubit quantum circuit simulated on commodity hardware via mathematical simplification.
The pattern. A common reflex when a problem looks expensive to compute is to throw bigger infrastructure at it. But for a large class of problems, the right mathematical trick can collapse the work into something a laptop can do.
What we built. A 60-qubit quantum circuit simulated on a laptop. The structure of the circuit gets collapsed into something small enough for commodity hardware to compute, then the answer is sampled one step at a time. The output bitstring is recovered via majority vote.
Why it matters. The interesting result isn’t the qubit count — it’s that the right mathematical instinct can replace a supercomputer. The same instinct (compress the representation, then compute) shows up across the everyday AI work we ship for customers.
If you have a workload that “feels like it needs more hardware,” we can often find a representation that doesn’t. That work shows up most often in retrieval, forecasting, and large-batch document processing.