Guillaume BalaineHourly
€160 / hour
Architecture reviews, code and prompt reviews, a second opinion before a big decision.
Minimum 2 hours
Freelance · 1–2 days a week · Remote worldwide · Paris (CET)
Your agent works in the demo. I make it work in production.
Part-time AI product engineering for startups: evals, typed outputs, routing, tool design, latency and cost. From an ex-CTO who built his own LLM agent harness in 2024 and now builds agents that ship code.
A senior AI engineer for your agent, one or two days a week.
Ex-CTO, hands-on. I harden LLM agents for production: evals on real traffic, schema-validated outputs, safe tools, routing, and cost you can predict. Without a full-time hire.
Since 2024
running my own LLM harness: typed outputs, routing, retry and repair, constrained tools
4 warehouses
SQL agent querying customer production data on BigQuery, Snowflake, Redshift, Databricks
Minutes
to ship a new agent tool instead of weeks, with config-driven tool definitions
Every day
building agents that plan, implement, review and ship software end to end
The problem
Getting an agent to a demo takes a week. Getting it to production takes judgment.
- 01Outputs break on edge cases, and you find out from customers.
- 02There are no evals, so every prompt or model change is a gamble.
- 03Tool calls invent parameters, and the agent loops or stalls.
- 04The token bill grows faster than usage, and latency hurts the product.
- 05The agent touches customer data in ways nobody has fully reasoned about.
Rates
Simple to start. Better with commitment.
Longer and larger engagements get a lower day rate. Treat these as starting points, not a final quote.
SprintMost teams start here
€1,200 / day
5–15 days, fixed scope. A production readiness audit: an eval suite on your real inputs, a failure-mode map and a ranked hardening plan.
Quoted as a fixed price up front
Retainer
€1,050 / day
1–2 days a week for 3 months or more. I own your agent and LLM layer alongside your team.
Fixed weekdays, 30 days' notice
Long-term
from €950 / day
6 months or more, or a larger scope across several teams. Let's shape it together.
Let's talk
Excl. VAT. Invoiced in EUR, USD or JPY. Pre-seed and need flexibility? Ask about part cash, part equity.
How it works
Measured before it's changed.
- Step 1
Intro call
30 minutes on your product, your agent and what's hurting. You'll know by the end whether I can help.
- Step 2
Measure
Evals and tracing on real traffic first, so every later change has a number attached to it.
- Step 3
Harden
Fix the failure modes that matter most: output contracts, tool design, routing, fallbacks.
- Step 4
Hand off or continue
Your team owns everything, documented. Keep me on a retainer only if it's paying for itself.
Selected work
Agents I've built, not just advised on.
An in-house LLM harness, built in 2024
Typed, schema-validated function definitions, task classification and routing, retry and repair on malformed outputs, constrained tool calling over a custom SSE protocol. It powered an analytics agent that generated correct SQL across four warehouse dialects and queried customer production databases safely.
A finance agent for CEOs, with bounded tools
Financial modeling, scenario planning, goal tracking and bank-invoice reconciliation. At its core, a graph-based modeling engine that uses iterative-deepening search to give the LLM bounded, context-aware calculation and lookup tools.
An agentic software-development system
Autonomous agents that plan, implement, review and ship changes end to end. The orchestration, eval and review patterns I bring to clients are ones I run every day.
See it on GitHubThe data foundations agents depend on
A real-time bidding platform at 1.5M requests a second and 30 ms P99, running online ML on ~100 TB of data a day. Profiling the hot paths cut infrastructure from $300k to $100k a month, the same discipline I apply to token cost and latency.
About
Hands-on, with a CTO's sense of what matters.
I'm Guillaume, a Paris-based engineer with 15 years in stream processing, real-time analytics, ML platforms and the products on top of them. I've been CTO three times, grown an engineering team to 12, and taken a company through SOC 2 Type II.
I work part time on purpose: the rest of my week goes into building my own agentic products, so the patterns I bring you are ones I'm using daily.
Working hours
Based in Paris (CET/CEST), remote worldwide, on site anywhere in Europe. Overlaps US East mornings and Japan late afternoons.
Languages
French and English, bilingual. Japanese, Italian and Spanish, intermediate.
TypeScript · Python · Rust · PostgreSQL · Kafka · Spark · Arrow / DataFusion · BigQuery · Snowflake · Cloudflare Workers · AWS · GCP · Kubernetes · OpenTelemetry
FAQ
Good questions to ask a freelancer.
- Is part-time enough to make a difference?
- For this kind of work, yes. Evals, harness design and architecture reward judgment more than hours. You get senior attention on the hardest layer without paying for a full-time hire you may not need yet.
- Are the rates negotiable?
- For longer or larger engagements, yes. That's what the tiers are for. Tell me the scope and the duration you have in mind and I'll send a proposal that fits.
- Do you work with teams outside France?
- Yes. I work remotely with teams anywhere and come on site anywhere in Europe. From Paris (CET), my day overlaps US East mornings and Japan late afternoons, and most of the work is async: written updates, pull requests, recorded walkthroughs. I work in English or French, and I'm happy to work with teams in Japan.
- Who owns the code?
- You do. Everything built for you is yours. General-purpose tools and know-how I bring in remain mine, and that's written into the contract up front.
- Which models and stacks?
- Model-agnostic: I work with the major hosted APIs and open models. Most of my work is in TypeScript and Python, with Rust where performance demands it.
Have an agent that needs to grow up?
Tell me what it does and where it breaks. I reply within two working days.
Or book a 30-min call directly