I build Excel and Google Sheets automation, Python pipelines, Claude and MCP agents and machine-learning models, and I prove each one on your real data with a measured before and after.
Every number below comes from a run you can replay: the test set, the transcript or the pull request is linked.
A support chatbot over 12 synthetic documents answered 20 of 30 test questions correctly. Only the pipeline changed (chunking, retrieval, dates, prompt). After: 29 of 30, outdated facts given as valid 3 → 0, 29 answers cite a source. Each version run twice, same scores.
Read the case study (56-s video)An agent skill for NextBrowser: the agent lets a password manager fill the login and reports one of eight fixed outcomes, stopping safely on 2-step codes, a locked vault or duplicate logins. Accepted after recorded runs on Microsoft, GitHub and Google; the recorded runs caught two bugs, fixed before the merge.
See the pull requestRead-only and sandboxed. On a 2,000-row synthetic workbook, Claude answered a business question in 30.5 s with 8 tool calls and saved an Excel report; 25 of 26 claims checked out; 81 automated tests.
See the MCP server demoOne click on 500 synthetic orders: 464 clean, 34 rejected with the reason, USD at live rates, summary and chart. 31 of 31 tests.
See the Sheets report demoEach demo opens in your browser. The live apps run on a free server and can take up to a minute to wake, so every page shows the recorded run first.
A forced network error retried after 2 s with backoff, a worker killed mid-job resumed from its checkpoint, every final answer must pass code checks before a human review; cancellation and cost per run. Every screenshot from a real run.
Open the workflow runnerPaste a site or upload PDFs: it answers only from that content, shows the exact passage it used, says when it doesn't know, and extracts the fields you name into a table. The same engine runs as a chat widget on your own site.
Open the chatbot See it as a widget on a siteDemo on synthetic data: four models tested per product, the most accurate kept, an order plan in whole cases, a budget optimizer over 300 demand scenarios per product.
Open the plannerA weekly roster that respects rest, contracts and leave, aims for a senior on every shift, and names the rule that blocks a shift instead of failing silently; type "Ahmed is off Thursday" and it re-plans.
Open the shift schedulerOn the sample cities 11–32% lower daily cost than a manual plan; Casablanca sample, 25 deliveries: 3 vans by hand, 2 vans optimized, about 40 km and $90 less per day. The map replays one recorded run.
The manual plan sends each van to the nearest stop that still fits; the optimizer plans all vans together under the same rules, on real roads.
Open the route optimizerVehicles and pedestrians tracked, drivable area and lanes, 29 frames per second on a laptop GPU.
YOLO11n, ByteTrack, YOLOP, ONNX Runtime
Watch the demoIn production for six months
Operators planned an industrial pipeline network by hand, a week ahead: which feeders run, where the flow goes, when the pipeline pumps or stops, with 12 storage tanks kept within safe limits and clients served around the clock.
I built a constraint-programming model of the whole network, about 10,800 decisions and 15,000 operating rules per run, with the pipeline's 25-hour transit delay built in. A simulator validates every schedule before it is used, projection alerts fire before a limit is hit, and the planning systems request new schedules through an API. In production for six months.
| Before | After | |
|---|---|---|
| 168-hour schedule | Hours of spreadsheet work | About 30 seconds on a normal day, under 5 minutes in complex scenarios |
| Operating rule violations | Found after the fact | Caught by the simulator before use |
| What-if (feeder outage) | Not possible | One API call |
Python, OR-Tools CP-SAT, FastAPI, Docker
Projects range from a one-day fix to a multi-week build.
Send me the file, the code or an example of the wrong answer through the platform where you found me, and I'll reply with the deliverable, the timeline and a fixed price.
