Ingestion
PDF, Word, Excel, AutoCAD. 40,800 distinct documents.
Independent AI engineer — Casablanca · on Paris time
An answer without its source is worth nothing.
The moment the stakes are contractual, a buyer doesn’t want a good answer: they want to verify it in three seconds. I build the systems where that’s possible — provenance carried from OCR all the way to the screen.
Projects from €8,000 · Day rate €500 · available now
Demo — tender file no. 42/2025 · 312 pages · FR/AR scanned
Query
What is the bid bond, and by what date?
Answer
The bid bond is set at 150,000.00 MAD, to be lodged no later than 14 November 2025, 10:00. The required certification is sector E — class 3.
3 values · 3 sources · 0 unsourced
Sources — click a citation
A reconstructed demo built from a typical file, values anonymised — this is not a client document. In production: 274,376 citations across 58 fields, 99.95% of them carrying the verbatim passage.
Measured in production — BidTender, since Sept. 2025
45,000files in production · 56 GB
274,376provenance citations
90.4%accuracy · hand-annotated set
1 mssearch over 113,000 vectors
Why RAG systems break in production
Three reasons, and none of them is the model.
- 01
Real documents are dirty
Scans, photocopies of scans, tables that exist only as images, structure that changes with every issuer. The pipeline that works on your ten test PDFs does not survive the eleventh.
- 02
Dense retrieval alone doesn’t discriminate
In a legal corpus, phrasing repeats. Two passages saying opposite things have near-identical embeddings — and what users actually search for, an article number, a date, an amount, is exactly what embeddings handle worst.
- 03
Without citations, the system is unusable
The moment the stakes are contractual, an unverifiable answer is worth nothing. And traceability isn’t bolted on at the end: it gets lost at chunking. It has to be carried from OCR to the screen.
The pipeline
— what carries provenance from OCR to the screenOCR routing
Page by page, bilingual. Only 14% of pages.
Chunking
By clause, not by window. 151,661 chunks.
Retrieval
pgvector + BM25, HNSW. 113,244 vectors · 1 ms.
Citation
The verbatim excerpt, kept. 99.95% of 274,376.
Interface
Verifiable in 3 seconds. 4 h → 40 min.
Ways to work together
- 01
Audit and scoping
A few days to establish why your retrieval is failing, what’s fixable, and what it costs. You leave with a costed plan — whether or not you hand me the build.
- 02
Production build
The full system: ingestion, retrieval, citations, interface. Shipped with an evaluation set that belongs to you, so you can judge quality without me.
- 03
Ongoing support
One to two days a month to evolve an existing system, arbitrate technical choices and hand knowledge over to your team.
Four references
See all →BidTender — multi-tenant SaaS
Citable RAG on documents that break pipelines
A Moroccan public tender file arrives as ten to thirty documents — PDF, Word, Excel, AutoCAD drawings — in French and Arabic, often scanned, up to 3,592 pages for a single contract. I built the system that reads them, with every extracted value tied back to the passage that produced it.
4 h → 40 minto triage a file · 10 → 30 files/month, same headcount
Leyton — consulting engagement
Real-time personalisation with a RAG persona pipeline
A cosmetics platform was recommending from frozen segments while ignoring all the text it already owned. Moved to per-person, real-time recommendation built on retrieval — and explainable by the business team.
+15%conversion rate · −20% analysis cycle
Chemin du Roi — industrial platform
Digital twin of an anaerobic digestion plant
Five heterogeneous real-time streams, five protocols, five notions of time — unified into one platform that plant engineers make decisions on every morning.
5 → 1protocols unified · OPC UA · Modbus · MQTT · SFTP · API
WardropAI · Wadrari
Two mobile products designed, built and shipped solo
What these two apps prove comes down to one word: I finish. A client handing over a document system fears receiving an impressive notebook and nothing that runs.
Shippednot a prototype · Kotlin · Expo · Supabase RLS
Got a corpus that fights back?
Describe your documents in three lines — volume, format, what you need out of them. In thirty minutes I’ll tell you whether it’s feasible, how, and what it costs. Reply within 24 hours on weekdays.