Notes on model collapse, synthetic data validation, and keeping AI training data quality high enough to bet a training run on — written by the Genovo Technologies team that builds Synthos.
July 8, 2026 · 6 min read
Model collapse quietly degrades models trained on synthetic or recursively generated data. Here is how it happens, how to detect it before training, and why AI training data quality is now a pre-training problem.
Oluwatosin Abioye Afolabi · Founder & CEO
Read the post →July 7, 2026 · 5 min read
Collapse is not a tabular-only disease. Why Synthos ingestion went multimodal — thirty-plus formats from parquet to WebDataset tars — and how validation itself follows.
John “Virus” · Machine Learning Engineer
Read the post →July 7, 2026 · 4 min read
How a Lagos-built deep-tech brand shows up in a global market: voice, color, credibility signals, and why we lead with engineering receipts instead of adjectives.
Osagie · Social Media & Design
Read the post →July 6, 2026 · 4 min read
The no-jargon tour: what model collapse is, what a Synthos validation actually gives you, what it costs, and how to run your first one today.
Osagie · Social Media & Design
Read the post →July 6, 2026 · 5 min read
What a dashboard shows when there is nothing to show reveals its values. How Synthos designs empty states, sample-data badges, and feature availability without faking anything.
Joy Ojochegbe · Product & Brand Design
Read the post →July 5, 2026 · 5 min read
Detection, containment, and honest status pages: how Synthos prepares for the bad day, and why impersonation tooling and audit trails are incident-response features in disguise.
Alayo Micheal · Cybersecurity Engineer
Read the post →July 5, 2026 · 5 min read
A quality claim you cannot independently check is just marketing. Why every Synthos certificate is publicly verifiable, expires in 90 days, and names exactly what was validated.
Adegbitẹ Ifeoluwa · Co-Founder & CTO
Read the post →July 4, 2026 · 5 min read
Dashboards are for humans; pipelines are where validation becomes habit. Designing the Synthos REST API, named scoped keys, webhooks, and the pip-installable SDK.
Chiebidolu “Chiboy” · Engineering Lead
Read the post →July 3, 2026 · 5 min read
A risk score tells you something is wrong; findings tell you where. How Synthos samples up to 5,000 rows to surface specific problematic records, and the honest limits of sampled inspection.
John “Virus” · Machine Learning Engineer
Read the post →July 2, 2026 · 5 min read
Real datasets ship as folders of shards, not single files. The design of dataset groups in Synthos — folder-native ingestion, group-level validation, and honest interim semantics.
Gasper Samuel · Product Manager & Engineer
Read the post →July 1, 2026 · 6 min read
How a small team ships fast on a platform customers bet training runs on: forward-compatible APIs, graceful feature degradation, typed contracts, and CI that earns its keep.
Chiebidolu “Chiboy” · Engineering Lead
Read the post →June 30, 2026 · 5 min read
A validation company cannot ship misleading charts. How Synthos picks chart colors with a computable process — lightness bands, CVD separation, contrast — instead of taste.
Ruby Cotterell · Product & Brand Design
Read the post →June 28, 2026 · 6 min read
A real incident: bot registrations flooding the user table with dot-trick emails. How we diagnosed it, layered CAPTCHA and rate limits, and what it taught us about abuse-driven threat modeling.
Alayo Micheal · Cybersecurity Engineer
Read the post →June 26, 2026 · 5 min read
Audit trails, retention honesty, and verifiable certificates: how compliance requirements become engineering specifications at Synthos instead of paperwork.
Iseoluwa Promise · Cybersecurity Engineer
Read the post →June 24, 2026 · 6 min read
How we replaced walls of bordered cards with hairlines, ambient depth, and a floating glass shell — and why a data-heavy product needs calm more than decoration.
Joy Ojochegbe · Product & Brand Design
Read the post →June 22, 2026 · 6 min read
A validation platform must treat every uploaded dataset as both precious cargo and a potential attack vector. How Synthos isolates, verifies, and protects customer data end to end.
Adegbitẹ Ifeoluwa · Co-Founder & CTO
Read the post →June 20, 2026 · 7 min read
The ML engineering inside a Synthos validation: stratified sampling, a 1M–500M parameter proxy cascade, curve fitting across scales, and confidence intervals we are willing to back with money.
John “Virus” · Machine Learning Engineer
Read the post →June 18, 2026 · 6 min read
What it takes to attach money to a model prediction: eligibility thresholds, human review, claims that pay out, and why a warranty changes the engineering culture behind it.
Gasper Samuel · Product Manager & Engineer
Read the post →June 15, 2026 · 5 min read
Teams do not skip data validation because they are lazy — they skip it because of how humans reason about invisible risk. What behavioral science says about why model collapse blindsides good engineers.
Oluwatosin Abioye Afolabi · Founder & CEO
Read the post →June 14, 2026 · 6 min read
An upload endpoint that accepts 500GB files in thirty formats is a pen-tester’s playground. How we attack our own ingestion path before anyone else does.
Iseoluwa Promise · Cybersecurity Engineer
Read the post →June 12, 2026 · 6 min read
Big datasets do not arrive as one tidy file. The engineering behind Synthos uploads: direct-to-storage signed URLs, resume after failure, multimodal format support, and whole-folder ingestion.
Chiebidolu “Chiboy” · Engineering Lead
Read the post →June 10, 2026 · 5 min read
Validation demand is bursty, tied to training cycles, and wildly uneven across teams. The product reasoning behind credit-based pricing — 25 for standard, 50 for express — and what it optimizes for.
Gasper Samuel · Product Manager & Engineer
Read the post →June 8, 2026 · 7 min read
Synthos trains 15–18 proxy models from 1M to 500M parameters per validation and extrapolates training outcomes with 90%+ accuracy. The research reasoning behind multi-scale cascade validation.
Oluwatosin Abioye Afolabi · Founder & CEO
Read the post →June 6, 2026 · 7 min read
Before any proxy model trains, Synthos screens corpora against known collapse signatures. A field guide to the statistical fingerprints of recursively generated data.
John “Virus” · Machine Learning Engineer
Read the post →June 5, 2026 · 6 min read
Synthos exists to certify other people’s data, so its own architecture has to be beyond reproach. How we apply zero-trust principles across tokens, roles, uploads, and the validation pipeline.
Adegbitẹ Ifeoluwa · Co-Founder & CTO
Read the post →June 3, 2026 · 7 min read
From upload to signed certificate in under 48 hours: how Synthos orchestrates sampling, 15–18 proxy training runs, extrapolation, and report generation in Go on AWS.
Chiebidolu “Chiboy” · Engineering Lead
Read the post →June 1, 2026 · 6 min read
Synthetic data now feeds most serious training runs, and model collapse is the tax nobody prices in. The founding story of Synthos, and why training data validation had to become its own discipline.
Oluwatosin Abioye Afolabi · Founder & CEO
Read the post →