Service
AI integration and automation for your product
I add AI to real products, not demos: assistants that answer from your own data, pipelines that sort and summarize incoming work, and agents that take a request and carry it to the finish. Each one is wired into your existing app, with limits on cost and a human in the loop where it matters.
What I build
Chat assistants that answer customers from your catalogue, documentation or account data. Triage that reads tickets, emails or forms and routes them with a summary. Generation pipelines for product content, images and video. Engineering agents, like devAgency, that turn a plain-language request into shipped code.
The stack
The Claude and OpenAI APIs, called from Node.js or Python back ends, streaming answers to a web or mobile front end. Retrieval over your own data so answers stay grounded, and structured outputs so the AI feeds your system instead of just producing text.
Built to be safe in production
AI features fail in ways normal code does not, so the guardrails are part of the build: rate limiting, daily cost caps, fallbacks to a human channel, and logs you can audit. Nothing an AI writes reaches your customers or your codebase without the checks you choose.
How a project runs
I start from one concrete task you or your team still do by hand, measure it, and automate that first. Once it works on real data, the scope widens. Billing is time and materials, and full ownership of the custom code transfers to you on final payment.
Projects in this area
Bilingual Shopify Storefront - Outdoor Living Brand
A custom Shopify storefront built from scratch for an Israeli outdoor-living brand: no template and no page builder, just a hand-written Online Store 2.0 Liquid theme, fully bilingual Hebrew/English with true RTL, and a 141-product catalogue migrated behind 209 SEO redirects. Every product ships in two art directions (a studio packshot and a Mediterranean lifestyle scene, 282 AI-generated images curated by hand), switchable site-wide from one header toggle that follows the visitor across cards, product pages, hero video and search. An AI shopping advisor answers sizing, material and composition questions against the live catalogue in both languages, streaming, behind durable rate limiting, a daily cost breaker and a WhatsApp fallback. Rounded out with an IS 5568 accessibility pass, structured data, predictive search and a performance budget. Built end to end, but the client put the launch on hold, so the storefront never opened to the public and nothing in it was ever orderable. What you see here is the finished build, not a shop.
Fixhos
AI ticket management platform for hospitality that automates guest request handling for hotels on Airbnb and Booking.com. Features multi-channel intake (WhatsApp, email, Booking, Airbnb), automated ticket classification and prioritization, SLA-based escalation, staff coordination via WhatsApp/email/voice calls, a marketplace of vetted technicians with quote approval, and a live operations dashboard.
Fashnable
AI virtual try-on for online stores: drop a flat garment photo and get a realistic video of an AI model wearing it. The app generates an on-model shot, with selectable model gender, age, body type, ethnicity and background, plus optional face and background references, then animates it, replacing costly photoshoots. Built as a full SaaS: invite-only auth, a credit ledger, asynchronous generation via Cloud Tasks, durable Firestore + Cloud Storage, and Polar billing with credit packs and subscriptions.
Brollwise
AI video editor for talking-head clips: drop a raw video and walk it through six steps, each paused for a human check, to a finished video with styled captions and b-roll. WhisperX transcribes it with word-level timing into an editable transcript you can export as .txt, dead air can be cut automatically, and captions come in four styles (Classic, MrBeast, Hormozi, Clean) with an optional karaoke mode that pops each word as it is spoken. Claude reads the transcript and your project context, picks the words worth highlighting and plans a cutaway every 5 to 8 seconds, using your own reference images where they fit and Flux-generated images where they don't, and any plan can be adjusted with a plain-English instruction. FFmpeg composites the final render. Hosted on Cloud Run behind invite-only Firebase sign-in, with direct browser uploads to Cloud Storage, each user's own API keys remembered in Firestore, and a running cost history per session. Started as a fork of the open-source ButterCut project.
Snapstems
AI stem separator as a paid SaaS: drop a track and get clean vocals, drums, bass and other stems, with 4-stem, 6-stem (adding guitar and piano), high-quality and karaoke presets, as mp3, wav or flac, played in the browser or downloaded as a zip. Audio uploads straight from the browser to Cloud Storage over resumable signed sessions, so large files bypass the server; the separation itself runs on Replicate's hosted Demucs, and a signed webhook finishes the job instead of a task queue, so the Cloud Run service sits idle while the model works. Credits live in a Firestore ledger, with 3 free splits after email verification, packs and monthly plans sold through Polar, refund handling and a reconcile tool. Own email-verification flow, a 10-language interface and an account and billing page. Python standard library only on the backend, no framework.
Hostaluego
Prospecting engine for a dealer that sells and leases copiers and telecom to French businesses and public buyers. In that trade the only real buying trigger is a contract coming to an end, so the app hunts timing, not copy: it pulls companies from the French government's open company registry and spots dated moves like a head-office transfer, reads public procurement data (DECP) for copier and telecom contracts nearing their term, and scores every lead out of 100, with the breakdown shown under each grade. Scores are recomputed every time they are read, weigh the kind of event and the distance to the sector's home département, and old or low-score leads age out so Prospection refills itself. Each qualified lead is checked on Google Places, and only the ones with a reachable French phone number are kept. On any lead, Claude writes a call brief, the "Déballe": why now, what is known with its source, an estimated monthly print volume and a ready phone opener, and it never mistakes a leasing company holding the contract for a competing dealer. Leads move into a built-in CRM with follow-ups, a daily call list, a planning view and a dashboard, where each contact carries its stated contract end (scored and planned) and the reason a lead was lost, and a call recorded on speakerphone is transcribed with Google Speech-to-Text and turned into a call report by Claude. Also: per-sector territories, a customer file import (CSV or Excel) that keeps existing clients out, an admin area with per-user daily lead limits and a read-only view-as mode, and a costs page that tracks every paid API call. Next.js on Firebase App Hosting with Firestore, a public landing page and invite-only sign-in, French UI.
Questions
- Which AI model do you use?
- It depends on the task. I mostly work with Anthropic's Claude and OpenAI's models, and I keep the integration simple enough to switch provider if pricing or quality changes.
- Is my data used to train the models?
- Business API traffic is handled under the provider's API terms, which differ from consumer chat apps. I check the current terms for your use case and keep sensitive data out of prompts when it is not needed.
- How do you keep AI costs under control?
- I build usage limits and a daily cost cap into the AI feature itself, and pick the smallest model that does the job well.
Want something like this?
Tell me what you are building, or what you are still doing by hand, and I will tell you how I would approach it. No commitment.