Economic intelligence,
for the institutions that need it.
Open forecasting infrastructure for banks, asset managers, and businesses. 25+ live data sources, regime-switching econometric models, and fine-tuned foundation LLMs — built from Nigeria, deployable for any emerging market.
See it work
A 40-second look at Asotele answering questions about Nigerian economic data — citing the actual source each time, and refusing cleanly when the data isn't there.
Video transcript & description
The video has no narration — every screen carries on-screen text, set against AceStep instrumental music. Sequence of frames:
- Title card: "Nigerian economic intelligence, grounded in citations."
- Live chat — MPR question: User asks "What is the current MPR?" Asotele answers "26.5% — cited to today's CBN briefing."
- NGX coverage: "Every NGX-listed company. Indexed and citable in seconds."
- Live chat — Dangote IPO question: User asks "Is there a pending Dangote IPO?" Asotele returns the actual NGX corporate-actions filing.
- Trust frame: "Asotele refuses to fabricate information. Cites every source."
- Live chat — FX prediction refusal: User asks "What will the FX rate be tomorrow?" Asotele refuses cleanly. Caption: "Asotele reports measured history, not predictions."
- Closing: "Asotele — Built in Nigeria. Open by design." Ends on the URL
asotele.apexgridapps.com.
Duration: ~40 seconds. No spoken audio.
Current indicators
Refreshed by the daily pipeline. Source files versioned in the project repo.
Daily briefing
Each morning Asotele delivers a Nigerian-context economic briefing — alerts, economic dashboard, news intelligence, CBN monetary-policy tone, Asotele Signals (AII + Sahm traffic lights), public-health signal, and council verdict — to the Asotele advisory committee. Briefing content is not published on the public site.
Two audiences. One engine.
The same Asotele forecast engine serves two distinct surfaces. Banks pay for the infrastructure; small businesses get the outputs.
For financial institutions
Banks, asset managers, treasury desks, credit risk teams. Forecast outputs surfaced via REST API, embedded inside Finovamax, or deployed inside your VPC.
- Daily regime classification on currency pairs
- 1/7/30-day FX, oil, equities forecasts with confidence intervals
- Multi-variable transmission analysis (VAR + Granger)
- Native integration with Finovamax banking software
- Open-source methodology — auditable for regulators
For businesses & founders
The graduate building a drone factory. The market trader hedging dollar exposure. The startup founder planning capex. Free access to the same forecast engine via a multilingual web chat, in English, Hausa, Yoruba, Igbo, and Nigerian Pidgin.
- Plain-language daily economic briefings
- Ask-anything Q&A grounded in the forecast engine
- FX, inflation, fuel, and sector outlooks for planning decisions
- Multilingual access — work in the language you actually think in
- Funded by Apex Grid's institutional revenue + grant partners
Launching Q4 2026 Get notified when it's live
Recent project updates
-
2026-06-16R5 regressed. The corpus was upstream, not the trainer.The R5 QLoRA round came back at 10.5% combined eval — a regression from R4a's 20.9%. The diagnosis took half an evening: the 59 new R5 training pairs all wrap citations in <sources>document/chunk_NNNN</sources> tags that don't appear in any of the 4,631 rows of R4a's existing corpus, and zero of them mention any of the 19 named signals the eval's must_cite axis scores against. The model was taught to cite the wrong thing. We patched the upstream extractor, added a style-filter floor on the promotion pipeline, and shipped a labelled-sentiment corpus that finally closes a gap the 2026-06-08 and 2026-06-10 Pidgin posts left open.
-
2026-06-14An advisory committee, an application form, and a private portalWe've opened applications to the Asotele advisory committee. The portal at asotele.apexgridapps.com/advisors is now live, gated by Cloudflare Access, and built around a deliberately light advisor role: async only, no calls, public attribution, honorarium available. The application form at /advisor-apply.html lives outside the gate and routes to a small queue we review by email. This post is the public record of why the committee exists, what advisors actually see when they sign in, and what we are not asking of them.
-
2026-06-10The Pidgin corpus is built. The number didn't move. Here's what we learned.Following the honest 0.460 macro-F1 reported in the 2026-06-08 Pidgin post, we built the BBC News Pidgin domain corpus the LINGUA Africa proposal scopes — 776 articles, 41,208 cleaned paragraphs, four years of coverage. We then ran the workstream-2 hypothesis end-to-end: AfriBERTa masked-LM continued pretraining on that corpus, then re-fine-tune the sentiment classifier, then test on AfriSenti. Result: +0.0009 macro-F1. Functionally zero. Adding class-weighted loss on top: +0.0052 best-case. The bottleneck isn't Pidgin representation — it's the structural 1.4% neutral class in AfriSenti pcm, and no amount of Pidgin domain text will move that wall. This is the pivot post: the right workstream-2 measurement target is a hand-labelled BBC Pidgin sentiment test set, not the AfriSenti held-out test. We've built the sample; the labels are next.