Retrieval-augmented generation
KnowledgeAgent
An internal chatbot that knows your company's actual policies and processes, and always shows its source.
- Your own documents
- Every answer cited
- No hallucinated policy
AI agents for
We build AI agents for repetitive operational work, grounded in your data, connected to your existing tools, and supervised when confidence drops.
Scroll
Hold toblast
- Practice
- Applied AI engineering
- Based in
- Zurich, Switzerland
- Data
- Azure Switzerland North
About
Systems that act rather than answer, evaluated continuously, handed over with the keys.
Fenro is an applied AI engineering practice founded by Alex Feng and Philip Rocki. The name is the two surnames, which is also roughly how the company works: there is no account layer between you and the people writing the code.
More about usFocused systems.
Measured execution.
GroundedEvaluatedBoundedSupervised
From workflow to outcome
The practice
Applied AI engineering,from audit to production.
What we do
We build AI into the workyou already do.
Fenro is an applied AI engineering practice. Most engagements open with a fixed-scope audit of where AI genuinely saves your team hours, and close with a system running inside the CRM, ERP or inbox you already use. Between those two points we design it, build it, evaluate it against your real cases, and hand it over with the keys.
Much of that work is bespoke. Three shapes recur often enough that we have the architecture, the evaluation set and the failure modes already mapped.
Agentic AI · Vision
DocumentAgent
An AI agent that reads your forms, extracts what matters, and fills your systems, so your team stops retyping and starts reviewing.
- Scans, photos, PDFs
- Confidence scoring
- Exceptions to a human
Agentic AI
IntakeAgent
An agent that reads what a customer actually needs and drafts a ready-to-send, accurate quote, before your team even opens the email.
- Unstructured enquiries
- Your pricing rules
- You approve, then send
From claim to proof
A solution is a claim.A system is the proof.
Selected work
Agentic AI · Client work
NationalKnowledgeAssistant
A retrieval-augmented chatbot grounded in Pro Senectute's own documentation rather than general knowledge, deployed nationally on Azure Switzerland North — so every office and every caller gets the same accurate, up-to-date answer regardless of region.
Built for
View our workSystem diagramRAG · Azure CH-N
- NationalRollout
- 26Regional organisations
- 3000+Documents indexed
- 100+Questions handled / month
- SwissHosted infrastructure
Industries
Where these systems run
The work looks much the same wherever documents, rules and inboxes pile up faster than people can get through them.
- 01
Finance
Reconciliation, reporting and controls all run on documents that someone has to read first. An agent pulls the figures out of statements and invoices, checks them against the ledger, and puts only the exceptions in front of a person.
- 02
Insurance
Claims and policy intake arrive as scans, photographs and free text. An agent extracts the structured fields, validates them against the policy and the schedule, and routes anything it is unsure about to an adjuster.
- 03
Logistics
Orders, delivery notes and customs paperwork move faster than the team reading them. An agent lifts the line items, flags mismatches against the original order, and keeps the exceptions queue short enough to clear.
- 04
Public sector
Entitlements and procedures vary by region and are updated constantly. A grounded assistant answers staff and public questions from the documents in force today, with a citation back to the source behind every answer.
- 05
NGOs
Programmes, eligibility rules and funding conditions sit scattered across years of documents. An agent gives coordinators one place to ask, and turns the questions it cannot answer into a documentation backlog worth fixing.
How we build
What makes it safe to run
A demo only has to work once. These four are what let a system keep working after it is handed over, on inputs nobody wrote a test for.
- 01
Grounded
Answers come from your own source material, and every one carries a citation back to the document it came from.
- 02
Evaluated
A held-out test set runs on every change, so a tweak to the prompt or the index cannot quietly make answers worse.
- 03
Bounded
Scoped to a task the system can do reliably, rather than open-ended autonomy that drifts once it meets real inputs.
- 04
Supervised
Low confidence goes to a person instead of a guess, with an audit trail behind every automated decision.
Solution lifecycle
From a real problem to a system running in your business.
Four stages, scoped up front, with something to look at every week.
Start here
Telluswhatis slowingyourteamdown.
A 30-minute call, no sales pressure. You don't need a finished spec — bring the problem and we'll tell you honestly where AI actually helps and whether we're the right fit.