🧭

Mara Lindqvist

👤 Role

Product Manager, AI tools

🟢 Status

Open to rolesAdvisory

📍 Location

Berlin · CET

🛠 Focus

EvalsAgent workflowsUnit economicsTrust & safety

🔗 Links

mara@example.com · LinkedIn · GitHub

👋

Eight years in developer tools, the last four on LLM products. I work on the parts that decide whether a model becomes a product: evals, failure handling, cost per task, and the handoff back to a human. Currently at Helioscope, and quietly open to what's next.

Most AI features don't fail at the model call. They fail at the step before it and the step after it — which is where I spend my time.

🧠 Skills

The recurring jobs I get hired to do, and how I do them.

Gallery List 6 items

📏

Evaluation design

Turning "it feels worse" into a graded set with a pass bar, so a model change becomes a decision instead of an argument.

Braintrust LangSmith

🩺

Failure-mode mapping

Finding where a model is confidently wrong, then designing around it: confirmations, fallbacks, a clean exit to a person.

Error taxonomies Session replay

💸

Unit economics

Cost and latency per completed task, tracked per feature, so pricing and model choice stop being guesswork.

Token accounting Routing

🔍

Discovery with users

Watching people do the work before automating it. The workflow around the feature is usually the real product.

Interviews Workflow maps

🧪

Prototyping

I build the throwaway version myself. A working prototype settles more debates in an afternoon than a spec does in a week.

Python n8n Supabase

🔐

Trust & governance

Data boundaries, retention, permission-aware retrieval, and the audit trail enterprise buyers ask about on call two.

SOC 2 EU AI Act

🚀 Selected work

Three projects where the interesting problem wasn't the model. Click to expand.

▶Support triage agentHelioscope · 2025–26

Problem: An autoresponder answered 60% of tickets and quietly broke the other 40% — the ones that mattered.
What shipped: A confidence threshold tied to real customer harm, an escalation path agents trusted, and a weekly review where failures became eval cases.
Result: Wrong-confident answers down 41%. Deflection held at 58%. CSAT up 0.6.
Role: Lead PM, team of six

OpenAI Zendesk Braintrust

▶Eval harness for a 40-person product orgHelioscope · 2026

Problem: Four teams shipped prompt changes on vibes. Regressions were found by customers, roughly a week late.
What shipped: A shared graded dataset per surface, a CI gate that runs on every prompt diff, and a scoreboard nobody has to maintain by hand.
Result: Regression escapes down from 9 per quarter to 1. Release confidence stopped being a personality trait.
Role: Lead PM

Python GitHub Actions LangSmith

▶Permission-aware retrievalNorthrail · 2024–25

Problem: Assistant search could surface documents the asker wasn't allowed to read. Two enterprise deals stalled on it.
What shipped: Permissions enforced at retrieval rather than after generation, plus an answer trail showing which documents were used.
Result: Both deals closed. Security review went from six weeks to nine days.
Role: Senior PM

Postgres pgvector Okta

📌 How I work

Write the eval before the feature.

Ship the boring fallback path first.

Measure cost per completed task, not per call.

Stop saying "AI-powered" in the UI.

📬 Contact

Open to senior PM roles on AI products, and to short advisory engagements when a team needs an eval practice stood up. Replies within two working days.

MAdd a comment…