AI Integration Engineer · UK · Booking now
Try it · paste a URL or describe your business
Already scoped an AI feature, or want to pick one off the menu of 50+? Either way — fixed price, shipped in 1–3 weeks, yours to keep (or I'll host it for you).
▸ Production, not demos
I built and run my own B2B SaaS. I know what AI work actually costs — and how it breaks.
▸ Start where you are
Roadmap if you're stuck. Sprint if you've scoped it. Either way, fixed price.
▸ You own it after
Code, prompts, AI agents — maintained by me or handed over so your team runs it without me.
Try it · paste a URL or describe your business
Paste any company's homepage URL. I'll read the page and propose one AI feature I'd build for that team — in about 8 seconds.
You'll get back
A one-paragraph read on what your team does
One AI feature I'd build for you
Which package it fits — Roadmap / Sprint / Build-out
Approach sketch — model, data, risks
Seven years as a software engineer across finance, regulated industries, analytics, and telecoms. I work in Python (FastAPI, Django, Flask) and TypeScript (Node, Next.js), and I've spent the last year specialising in AI features, agents, and workflows that ship into production — not demos.
Most of my clients don't have a dedicated AI team. They're law firms drowning in document review, agencies running outbound by hand, e-commerce ops teams triaging support tickets, accountants categorising invoices, founders shipping their first AI feature. They know AI should help. They don't know where to start. That's the gap I sit in — translating “we should do something with AI” into a working tool their team uses on Monday.
I also built and run Filemender, a B2B SaaS for post-production teams, end to end and solo. The growth stack you can read about above runs my marketing function. I know what production AI work actually costs, where it breaks, and how to engineer around the failure modes — because I've hit them in my own product.
I work remotely across European business hours, take on one client at a time, and I'd rather turn down work than over-promise on a timeline. If we're a fit, I'll tell you. If we're not, I'll point you at someone who is.
Models & agents I build with · default Claude
Claude
Anthropic
▸ DEFAULT
GPT
OpenAI
▸ BENCHMARK
Gemini
▸ BENCHMARK
Grok
xAI
▸ BENCHMARK
Llama
Meta · Open-src
▸ ON-PREM
DeepSeek
DeepSeek · OSS
▸ ON-PREM
Cursor
Coding agent
▸ AGENT
ClickUp
Workflow agent
▸ AGENT
Claude is the default for production work (evals, long-context, tool use). I benchmark against GPT, Gemini, and Grok per project, and route to open-source Llama or DeepSeek for on-prem or cost-sensitive workloads. Day-to-day I work inside Cursor. Model choice is part of the deliverable, not an assumption.
Live · 8 scheduled Claude agents shipping into one business · looped preview
I built and operate Filemender, a media file validation and repair tool used by post-production studios and agencies. The interesting part isn't the product — it's the AI agents and workflows running on top of it.
Eight scheduled Claude agents handle SEO content, lead research, cold email drafting, LinkedIn posting, and Twitter monitoring. The whole marketing function runs as engineering — and the case study walks through what I got right, what I got wrong, and what I'd build differently for a client.
Case study · ~12 min read
Filemender — eight Claude agents running growth
Architecture, costs, what broke, and what I'd build differently for a client.
Scheduled Claude agents in production, each with a single narrow job
Cost reduction after migrating routine work from Opus to Haiku
Fabricated leads after adding the verification layer
What a user said
“FileMender saved us from a disastrous broadcast rejection on a Friday night. It fixed the loudness levels automatically in seconds.”
Post-production lead
Short engagements where I shipped a single AI feature end-to-end. Scope pinned, price pinned, handed over with evals.
Document review · 10 days
Replaced a paralegal workflow that took 3 hours per contract. Claude reads the PDF, extracts 18 named clauses into a structured review table, and flags anything unusual. Now runs in under 90 seconds per doc.
Classification · 10 days
Replaced a hand-written rules engine with a two-model classifier: Haiku for cheap triage, Sonnet for ambiguous cases. Cut ops time per lead from 4 min to 18 seconds.
Agents · ongoing
Daily agent that researches ICP accounts, writes a first-draft cold email, and queues it for the founder's review. Runs for <£30/mo of model spend.
More case studies on request · Ask to see one ↓
Most AI work that fails in real businesses fails for the same reasons: no clear scope, no evals, runaway costs, no handover. I package the work to remove those failure modes upfront — and I'll start with a roadmap if you don't yet know what to build.
▸ Every Sprint & Build-out ships with the learning kit so your team can run it without me
▸ See the full AI Agent Menu (PDF) — 50+ agents I can ship, tagged by package and time
AI Roadmap
1 week · ship one small agent
Best for: you want to see AI working in your business this week. Pick a small agent off the menu and I'll ship it. If you genuinely don't know what you need yet, swap it for a strategy plan instead.
Roadmap examples from the menu
AI Sprint
2 weeks · 1 medium OR 2 small
Best for: a specific AI agent you've scoped, or a pair of smaller ones. Most of the menu lives here — contract clause extraction, lead qualifier, support triage, CRM enrichment, that kind of thing.
Sprint examples from the menu
AI Build-out
3–4 weeks · 2 medium OR 1 large
Best for: a connected multi-agent setup or one ambitious agent. Two Medium agents bundled, OR one Large from the menu (e.g. account health scoring, anomaly detection). Narrow and deep, not sprawling.
Build-out examples from the menu
▸ Care subscription · £50/mo per agent · most clients pick this
I host the agent on my standardised stack, monitor it 24/7, and fix bugs free. New requests are quoted separately as a mini-Sprint. Cancel any month. Best for teams without an in-house tech function — which is most of them.
Every Sprint and Build-out ships with a full handover kit — six concrete artifacts that document everything I built, the decisions behind it, and how to run it. Whether your team runs the agent (Self-host) or I run it for you (Care), the IP, the code, and the prompts are yours from day one.
Included with every Sprint & Build-out · yours regardless of how it's hosted
▸ 01
30-minute Loom recorded during handover. Every decision explained — model choice, routing, fallbacks, why this and not that.
▸ 02
Every prompt in the system with comments on why it's structured that way. Lives in your GitHub if you Self-host, or your shared docs if I run it on Care. When it drifts, you know which knob to turn.
▸ 03
Step-by-step for the day-to-day ops: re-triggering a job, checking spend, rolling back a prompt, adding a new case. Yours to follow if you Self-host; mine to follow if you've added Care.
▸ 04
What breaks, what the signal looks like, what to do. Every hallucination class and edge case hit during build — documented, not hidden.
▸ 05
The test harness lives in your repo. Change a prompt or swap a model? Re-run in one command, see the delta, ship with confidence.
▸ 06
14 days on a Roadmap, 30 on a Sprint, 60 on a Build-out. Anything that breaks the agent doing what the SOW said — I fix it on me. Care extends this indefinitely (£50/mo per agent).
The answers most people want before they'll spend 15 minutes with a freelancer. If yours isn't here, just ask it on the call.
No — that's exactly what the AI Roadmap is for. Or, if anything in the Agent Menu (harrybenham.dev/menu) caught your eye, we can ship that as the Roadmap deliverable instead. 100% credited toward a Sprint if you proceed.
Most agents on the menu run on £20–£100/mo of API + hosting. Per-feature spend caps go into every Sprint so a runaway loop can't run away with your card. You get a costed estimate per use case on the Roadmap.
Default for most clients is Care (£50/month per agent): I host the agent on my standardised stack, monitor 24/7, fix bugs free, you never log into a hosting dashboard. New requests are quoted separately as a mini-Sprint. Cancel any month. If you have an in-house tech team and prefer Self-host, that's free — you own everything from day 30.
Scope is locked in writing on Day 0 — what's in, what's out, the eval set, success criteria. New features mid-build get quoted as a small add-on Sprint, never billed surprise-style. If I under-quoted, that's my problem, not yours.
The single most important question — which is why evals are non-negotiable in every package. I build a test suite covering the critical paths. For high-stakes outputs (legal, financial, customer-facing) I add a verification layer or a human-in-the-loop step.
Almost never. Most menu agents run in standard channels (web chat, email, WhatsApp, Slack, your CRM via API) and don't touch your product code. If something does need deeper integration, I work alongside your engineers — you stay in control of your codebase.
That's the most common case and exactly what Care is built for. £50/month per agent, I host and monitor it, you never touch infrastructure. New work gets quoted separately so there are no surprises. The agent just works — same way you don't manage your own email server.
Yes on both. Happy to sign an NDA before discovery. Client data never leaves your infrastructure unless you explicitly want it to. I route model calls through enterprise endpoints with zero data retention where requested. UK GDPR DPA available on request.
Fifteen minutes, no prep needed. Tell me what you're trying to build (or that you don't know yet), and I'll tell you whether I'm a fit, what it would cost, and how long it would take.
No slide deck
Just a conversation. Camera optional.
No hard sell
If we’re not a fit, I’ll say so on the call.
No follow-up spam
One email after with notes. Then it’s on you.
