Find your agent expert

Tel Aviv, Israel

I get called when an agent system works in the demo and fails in the field, and nobody in the room can explain the gap. How I work through it: Build the evaluation suite first. You cannot fix what you are not measuring, and almost nobody arrives with one. Reproduce the failures on real traffic, not on the happy path someone recorded for the board. Rank the fixes by cost. Two prompt changes and a retrieval fix usually beat the retraining that was proposed. Leave the harness behind, so the next regression is caught by your team rather than by a customer. Four enterprise deployments remediated this way. Former research engineer, independent since 2023, one slot roughly every two months.

Istanbul, Turkey

Turkish-language voice agents, a space where most off-the-shelf tooling still falls over somewhere in the first sentence. I run my own transcription evaluation instead of trusting a vendor benchmark, on a Turkish speech dataset I keep adding to. Half this job is telling people voice is the wrong channel for what they described. The other half is making it work when it is the right one. In production: appointment booking across 30 clinics in Istanbul and Ankara, and a collections calling agent for a consumer finance company that had to pass a legal review first.

Bogotá, Colombia

I collect the data other people say cannot be collected: public tenders, distributor price lists, marketplace inventory across Latin America. Solo, fast, short engagements with a clear finish line, and a handover a non-technical team can keep running once I am gone. A daily price monitor across 14 Colombian and Mexican retail sites and a tender watcher for a construction group are both still going, three years later. Mid-contract right now and not taking new work until it ends.

Tunis, Tunisia

Most of what I do is repair work. A team buys into a demo, the demo does not survive contact with real users, and I get called a quarter later to find out why. Nine times out of ten the model was never the problem. The tools it was handed were. So I write the MCP servers myself, typed and tested, and only then design the agent graph on top. Last year that meant turning a dead rollout at a French insurer into a nine-agent system that now runs unattended. Three of my banking-adjacent MCP servers are open source, and I speak at Francophone and Arabic-language developer events when the travel works out.

Dakar, Senegal

WhatsApp-first agents, because in this market that is where the customer already is and nobody is downloading your app. Deliberately simple, cheap to run, sized for a few hundred conversations a day rather than a few million. Two-week fixed-scope projects, priced before I start: an ordering agent for a Dakar grocery service, lead capture for three West African insurance brokers.

Doha, Qatar

Market intelligence pipelines for construction, energy and trading firms across the Gulf. The value is never the summary at the end, it is having read everything before your competitor did. Sources I handle without complaining: Tender portals in six GCC countries, several with no working search Regional news in Arabic, Farsi and English, all in one pipeline PDFs that are really just photographs of paper A commodity digest I built lands in a client inbox at six every morning and has not missed a day in two years. New work from next month.

Stockholm, Sweden

Most of the agent systems I get asked to review should be three cron jobs and a queue. I build the plumbing: tool servers, memory stores, and the evaluation harness that tells you whether last week made things worse. Small and boring by preference, and I will happily argue you out of a multi-agent design you do not need. An agent platform I built for a Nordic bank has been in production eighteen months. Two Series B startups run on my MCP layer. Free again from next quarter.

Milan, Italy

Ten years of ERP integration work, then agents got good enough to replace most of the glue code I used to write by hand. 2014 to 2019: integration consultant for Lombardy manufacturers, mostly SAP and a lot of legacy MES. 2020 to 2023: automation lead at a B2B distributor, 200 people, order intake down from four hours a day to twenty minutes. Since 2024: independent, building order and reporting agents for industrial suppliers who still run on spreadsheets and phone calls.

Dublin, Ireland

A revenue-operations studio: I wire agents into the parts of a sales org that leak money. Eight people, mostly ex-RevOps rather than ex-engineering, which shows in how we scope. Every engagement runs the same way: Two weeks reading your CRM instead of your brief. We come back with where the pipeline leaks and what it costs. Fix the data model. No agent survives a CRM where half the deals have no owner. Ship the automations that pay for themselves first, usually pipeline hygiene and outbound research. Hand over the reporting and get out of the way. Twenty-plus SaaS companies on HubSpot and Salesforce so far, including an outbound research agent that replaced a five-person SDR pod and quarterly board reporting with no analyst in the loop.

Accra, Ghana

Focus Support agents for fintechs and mobile-money operators across West Africa. Languages in production English, French, Twi and Pidgin, because that is what customers actually write in. What I care about Escalation rules and tone. An agent that invents a policy costs more than the queue it replaced. Running now 12,000 tickets a month at a Ghanaian lending app, first response under two minutes, prompts maintained by my team rather than the client's.

Hamburg, Germany

We're six people in Hamburg, working almost entirely on logistics and maritime. We take the projects where the data is the hard part: EDI files nobody documented, port APIs that answer in XML from 2003, forty years of shipping records in a system whose vendor no longer exists. The agent layer on top is usually the easy week. Two of our four founders came from terminal operations rather than software, which is why we argue about berth windows before we argue about models. Currently watching 3,000 shipments a day for a North Sea carrier and 90 retired Excel macros at a customs broker.

Nairobi, Kenya

Agents for East African SMEs: schools, clinics, logistics operators, anyone drowning in WhatsApp threads and paper forms. I design in this order: Works on a cheap phone with two bars of signal Works when the person who asked for it is on leave Looks nice, if there is budget left over Enrolment and fee-reminder agents for four Nairobi schools, dispatch coordination for a regional courier. I also train one person on your side to own the thing before I invoice the final milestone, which has never once been the person who hired me.

Copenhagen, Denmark

Operations automation for Nordic mid-market companies. Copenhagen, small team, mostly manufacturing and distribution. We run a strict two-phase model: Phase one: map the process end to end and agree, in writing, what should not be automated. Clients try to skip this phase and it is the one that saves the project. Phase two: build only what is left, which tends to be about 40 percent of the original brief. Order-to-cash for a Danish industrial distributor, and a service desk agent now absorbing 60 percent of internal IT requests. Next availability opens in a few weeks.

Kyoto, Japan

Retrieval over Japanese-language corpora, which behaves nothing like the English benchmarks everyone quotes. The hard parts Tokenisation, product codes that mix kanji and katakana, and vertical-text scans of manuals written before anyone thought about search. I fine-tune my own rerankers when the general ones give up. Who hires me Manufacturers and law firms sitting on decades of internal documents, with a very low tolerance for a confident wrong answer. Availability One project at a time. The next slot opens in a few weeks. I also write an occasional newsletter on Japanese RAG evaluation.

Pune, India

D2C back-office work: catalogue updates, competitor pricing, returns triage, marketplace reconciliation. Shopify, Amazon Seller Central, and three Indian marketplaces with no real API to speak of. Good fit: Founders who already know which spreadsheet is eating their week Catalogues in the thousands of SKUs across several channels, like the 18,000 I currently maintain for one brand Teams who want to own the thing afterwards Bad fit: anyone looking for a storefront rebuild. I work behind the storefront, not on it. No account managers in between either, you get me.

Mexico City, Mexico

I work with Latin American B2B teams on the gap between marketing and sales, which is where the deals quietly go to die. The order is always the same: rebuild the CRM first, then add agents. Skipping the first half is how a company ends up with an expensive assistant enriching garbage. Recent work includes lead routing for a 60-rep logistics sales team and an outbound sequencing agent producing around 200 qualified conversations a quarter. I have sold in Mexico, Colombia and Chile myself, so I know which parts of an imported US playbook to throw away.

Toronto, Canada

We're fourteen people in Toronto, on the long regulated projects: healthcare, insurance, public sector. Why clients pick us We arrive with our own evaluation and audit tooling. In these industries, "it seems to work" is not something you can put in front of a regulator. Recent A claims-triage agent reading 50,000 documents a month for a national insurer. A clinical knowledge base where every sentence traces back to a source document. One rule Two-week discovery before any code, including for clients who have already chosen their architecture.

Berlin, Germany

I work at the far end of the problem: when prompting stops paying off and the model itself has to change. Domain-adapted models for legal and pharmaceutical clients, two or three engagements a year, almost always with teams who already have something working and need it twice as good. I also advise three AI infrastructure companies. I book roughly six months out, so the honest answer to most enquiries is next year.

Dubai, United Arab Emirates

Outbound and qualification agents for real estate, wealth management and professional services in the Gulf. Ten years in regional sales before automation, which is where the compliance instincts come from. Best fit: high inbound volume, three languages in the same thread, and a sales team that will actually follow up on what the agent qualifies. Not a fit: cold outreach at volume into markets I do not know. Cheaper people will do that for you, and worse. Currently: a Dubai brokerage taking 1,200 inbound leads a month through one WhatsApp agent covering Arabic, English and Farsi.

Warsaw, Poland

I get data out of systems that were never meant to give it up: ageing ERPs, PDF exports, portals with a login and no API. I treat scraping as an engineering problem. Selectors break, sites get redesigned on a Friday, and a script somebody runs by hand is not a pipeline. Every handover includes: A runbook a junior can follow Alerting on silent failures, which are the expensive kind A written list of what will break first and roughly when Running today: nightly ingest of 40 supplier catalogues for a Polish distributor, and 15 years of scanned invoices turned into rows you can query.

Cape Town, South Africa

I work with small teams and nonprofits that need automation, do not have an engineer, and are not about to hire one. That single constraint decides everything else. Whatever I build has to be something you can open, understand and change three months after I am gone, which rules out most of the clever options and all of the fashionable ones. Intake and case tracking for three Cape Town nonprofits, volunteer scheduling to retire a shared spreadsheet: that is the shape of the work. I also run a free monthly workshop on agents for non-technical founders. Come to that first if you are not sure you need to pay anyone.

Austin, United States

Model Retainer, not projects. The interesting problems only surface in month three. Clients A handful of eight and nine-figure DTC brands, most of them with us over a year. What we run Merchandising agents repricing 60,000 SKUs daily, and a post-purchase support agent holding a 4 percent escalation rate. Not for us One-off builds, and anything where nobody in the room owns the P&L. Availability Taking bookings for next quarter.

Bengaluru, India

Six years on search infrastructure before agents existed as a job title. It shows in how I think about recall, latency, and what a wrong ranking actually costs. Before: query understanding and ranking at tens of millions of documents, the scale where the vector search everyone starts with quietly stops working. Now: retrieval for banks and large enterprises. A knowledge platform serving 8,000 internal users at an Indian bank. One rebuild that moved a RAG system from the right answer 40 percent of the time to 91. On the side: advising two AI startups on retrieval architecture, mostly talking them out of the fine-tuning they do not need yet.

Lisbon, Portugal

The unglamorous half of e-commerce for Iberian and Brazilian retailers: stock, returns, chasing suppliers, and margin numbers that arrive before the month is over. How it starts A week finding where the money is actually going. It is rarely where I think. Twice now the answer turned out to be freight rather than discounting. What usually gets built Inventory reconciliation across warehouses and marketplaces, and the weekly margin report that used to occupy an analyst for two days. Capacity Two clients at a time, no more. Next opening in a few weeks.

Oslo, Norway

Typed, tested MCP servers over internal systems, which is very nearly all I do A full tool layer over a Norwegian energy company stack, two years in production Contributor to several open-source MCP implementations One standing opinion: most agent failures are tool failures in a costume, and I keep the incident reports that say so Fully booked this year. Worth a message if you are planning something for after that.

Melbourne, Australia

We're five people in Melbourne, serving Australian professional services firms: accountants, agencies, recruiters. Every client starts with the same fixed-price discovery sprint: Week one, we sit with the people doing the work being automated. No workshops, no whiteboard sessions. Week two, a working first version in front of real users. It will be ugly. Week three onward, we iterate with you instead of disappearing into a spec document. Built this way: client onboarding for a 300-person accounting practice, and candidate screening for two recruitment firms that had both been burned by a bigger agency first.

Amsterdam, Netherlands

I replace the weekly reporting ritual with agents that write the analysis and flag only what moved. I came out of a data team spending 60 percent of its hours on recurring reports and have been undoing that pattern ever since, most recently for a 40-store Dutch retail chain and an anomaly watcher sitting on a Snowflake warehouse. I will also tell you when a plain dashboard is the better answer, which costs me work and saves you money.

São Paulo, Brazil

I automate the operations side of Brazilian e-commerce, including the parts that only exist in Brazil and that nobody from outside wants to touch. Marketplace rules that change without notice, nota fiscal, seven carriers with seven different ideas of what an address is. I have already walked into those walls, which is most of what you are actually paying for. Typical client is a mid-size seller doing 5,000 orders a month and up. Order and invoice sync across Mercado Livre, Shopee and their own store, plus a freight quoting agent that retired a spreadsheet three people were editing at once.

London, United Kingdom

Agent systems for financial services, where the review process is longer than the build and everyone involved has made peace with that. Who we are Twelve people in London. Half came from compliance or risk rather than engineering, which is unusual and turns out to matter more than anything else on this page. Recent A research assistant now in the hands of 2,000 analysts at a European asset manager. What ships with every system Full audit trail, model-risk documentation, and a red-team report we did not write ourselves. Timelines Six to nine months. Anyone quoting you six weeks has not read your policy.

Cluj-Napoca, Romania

What No-code tooling with agents on top, for small European teams who need something automated by the end of the month rather than the end of the quarter. Rate Flat weekly, no long contracts, stop whenever you want. Track record Over 40 short engagements with agencies and small SaaS companies. Onboarding and invoicing for a German staffing firm is the one that ran longest. Honesty clause When no-code stops being the right answer, I say so and help you hire the engineer instead.

Seoul, South Korea

Korean-language voice agents for commerce and customer service, tuned on real call recordings rather than clean read speech. Latency is a product feature. An agent that answers correctly in four seconds has already lost the caller. My order-status agent for a Korean home shopping network runs on a sub-second pipeline over a fine-tuned speech stack, and handles the callers who start talking before the greeting finishes. On a long engagement through the end of the year, so not taking new work.

Prague, Czechia

Four engineers in Prague. No sales team, no deck, most work arrives as a referral from someone we finished with a year ago. We build the unglamorous internal tooling agents need before they are useful to anybody: tool servers, permissions, the audit log someone will ask for in month six. An internal platform we built for a Czech telco carries 400 daily users, and our MCP servers for Jira, Confluence and GitLab are open source. We would rather fix an existing broken system than start from nothing. Greenfield is more fun and less useful.

Lyon, France

Voice and chat agents for clinics, hotels, and anyone whose phone rings more often than the staff can answer. 400 calls a week booked for a group of dental practices A bilingual front desk agent for a hotel group in the Alps, running through two ski seasons The escalation script written before the first prompt, every single time Getting a human on the line treated as a feature, not a failure state I worked front desk myself for three years before any of this, which is the only reason I know what a caller sounds like ten seconds before they hang up.

Ho Chi Minh City, Vietnam

Price and stock tracking across Shopee, Lazada and TikTok Shop A bulk listing agent pushing 5,000 product updates a day Competitor monitoring for sellers who are live on all three platforms at once High volume, low ceremony. I have run this playbook for dozens of Southeast Asian sellers, so I know where it breaks and I will tell you before we start rather than after. One to three week sprints.

Singapore

Ten people between Singapore and Kuala Lumpur, building agent systems for banks, insurers and government agencies across Southeast Asia. Why local matters here Data residency, four official languages in one support queue, and regulators who will ask what the model was trained on. We grew up in this environment rather than reading a summary of it. Selected work Onboarding and KYC document agents for two regional banks. A citizen services assistant answering in English, Malay, Mandarin and Tamil. Every project Ends with a knowledge transfer phase. If your team cannot run the system without us, we have not finished.

Kyiv, Ukraine

Six years in data journalism before going independent, which is why sourcing is not negotiable here: if an agent cannot show where a number came from, it does not ship. What that looks like in practice: Monitoring pipelines that stay silent until something has genuinely changed Competitive intelligence for a European logistics group, running two years now A regulatory change watcher covering 11 jurisdictions in three languages Available again next month.

Wellington, New Zealand

Half my engagements end with a smaller system than the client asked for. I count those as the wins. I advise engineering leaders on whether to build agents at all, and then on how not to overbuild them. Architecture reviews for more than 20 teams across APAC, three multi-agent projects talked out of existence that should have been scheduled scripts, and a long-standing suspicion of any diagram with more than five boxes in it. On sabbatical until later this year.

Amman, Jordan

Who I work with Small businesses across the Levant: 15 retailers, clinics and law offices so far, between Amman and Beirut. The part most vendors skip Dialect. Customers write in Levantine Arabic and get answered in Modern Standard, which reads like a letter from a ministry. I tune for the way people actually type. What makes it affordable An Arabic prompt library I maintain and reuse, so the tenth client pays a fraction of what the first one did. Starting New engagements from next month.

Barcelona, Spain

We're a Barcelona studio working with European fashion and lifestyle brands on the operations behind the storefront. Six people, half of them designers, which is why our internal tools look better than internal tools usually do. That sounds cosmetic until you watch a warehouse team quietly refuse to use something ugly. Recent: a returns and exchange agent for a brand shipping to 14 countries, and a product content pipeline writing listings in five languages. Retainer clients get a shared roadmap they can reorder themselves, which they do, constantly.

Taipei, Taiwan

Chinese-language agent systems for hardware and semiconductor companies, where the documentation is dense, half the vocabulary is internal, and the tolerance for a wrong number is close to zero. Two things I end up saying on most projects: Fine-tune only when it earns its keep. A better retrieval setup usually does the same job for a tenth of the cost. Traditional and simplified are not a rendering detail. They change what the retriever finds. Current work includes an engineering knowledge assistant for a Taiwanese component manufacturer and retrieval evaluation across both scripts. Three engagements at a time, maximum.