Case studies
Different industries, different objects, the same guarantees: the
decision is traceable, the record cannot be quietly edited, and a
person remains accountable for it.
01Banking - credit & lending
Credit appraisal decision support
We worked with a Tanzanian bank to improve its loan origination
process. Previously, every facility began as a folder of document
scans - certificates of incorporation, tax clearances, bank
statements, valuations - re-typed into forms by hand. Forged
documents could slip past manual review, duplicate and conflicting
data was common, and getting a full picture of the borrower - a true
valuation of collateral, a missing document - meant days of
back-and-forth with applicants.
What we deployed
- An ontology of the lending file: applicant, case, document checklist, captured fact, finding, appraisal.
- Every document read against a fixed schema for its type, with forgery checks per document type and an agent that verifies filings against the corporate registry - attaching what it found as evidence.
- Alternative data where the file is thin: collateral valued from market data rather than the applicant’s word, and statement transactions grouped by counterparty so the officer sees who the business actually trades with.
- Data lineage built in: every figure on the generated appraisal links back to its source, and every action in the system leaves an audit trail.
Results
- Loan appraisal process time reduced by 80%
- Forgery checks that previously depended on an officer’s eye now run on every document, every time
- One clean record per applicant and fact - duplicate and conflicting data eliminated at the source
02Investment research - tech & semiconductors
Self updating supply chain ontology
We built a research system for an investment firm covering tech and
semiconductors. Previously, high-quality information - paid
research, podcasts, X threads, a large internal archive - arrived
faster than it could be read. LLM summarization didn’t help:
summaries flattened the analysis, hallucinated claims were quite
common, and the models missed the indirect beneficiaries an article
implies but never names. Additionally, theses weren’t
continually and automatically updated as new information came to
light - a view written six months earlier stayed in place even
after new information had invalidated it.
What we deployed
- A bi-temporal knowledge graph of the semis value chain: every fact typed, anchored to its publication date, and carrying provenance back to the exact source passage. Contradictions close a fact’s validity window rather than deleting it - the system can answer “what did I believe in January, and why.”
- A reasoning layer with structural bias controls: a mandatory bear-case pass on every thesis, invalidation triggers written before conviction, and no visibility into positions - views are formed blind.
- A layered evaluation harness built around a hard constraint: models already “know” historical outcomes, so backtests grade the derivation (grounded mechanism chains, cited evidence), while true judgment is scored only on post-training-cutoff windows and a live ledger that records every call with its price at the moment it was made.
Results
- 96% of surfaced claims trace to a verifiable source passage; extraction, retrieval, and reasoning are each regression-tested against frozen gold sets
- Second-order calls - names the source articles never mentioned - outperformed a “buy the names in the article” baseline by 3.1% over 12 weeks in the first scored quarter