How Tracent Operates: Building an African AI-Native Company from Day One

The Naira is at ₦1,500 to the dollar this morning. Okra, once Nigeria's open-banking darling, has been wound down for over a year. The African deep-tech graveyard is well populated, and most of those companies did not die because their product was wrong. They died because their operations were the operations of a 2017 startup running in a 2026 economy. Burn rate set by US salary expectations. Hierarchies modelled on Silicon Valley org charts. Internal tooling treated as durable infrastructure when it should have been disposable. AI bolted onto workflows designed for human information routing.

We are building Tracent Technologies, an Agent-to-Transaction Middleware that makes African payment rails callable by AI agents under sovereign, NDPA-compliant rules. The customer-facing product is one bet. The way we operate is another. This post is the public commitment to the second bet, because the operating model is itself a moat, and because saying it publicly forces us to keep doing it.

The five loops

Tracent is structured as five named operating loops, not as a hierarchy. Each loop has a sensor layer (where information enters), a policy layer (rules about what to do), a tool layer (deterministic capabilities), a quality gate (human or automated review), and a learning mechanism (how the loop improves itself over time). Each loop has exactly one Directly Responsible Individual.

The Sales and Partnership Loop captures every inbound enquiry, every BD conversation, every partner interaction. Recordings drop into a private repository, get transcribed by self-hosted Whisper, and feed an agent that drafts follow-ups. The founder approves before anything goes out. The metric is conversion at each stage from inbound to paid, and the loop improves by reviewing closed-won and closed-lost conversations against the positioning that was used.

The Customer Health Loop watches login frequency, gateway latency, error rates, payment failures, and one-question NPS responses. A customer is flagged at-risk if they have not logged in for 14 days, or their tool error rate is above five percent, or a payment has failed and stayed unresolved for seven days. Founder attention triggers within 24 hours for top-revenue customers; the agent drafts outreach for the long tail.

The Compliance and Audit Loop is the most strategically important of the five. It is both the regulatory moat (NDPA registration as a Data Controller of Major Importance is mandatory before we serve real customers) and the customer-facing differentiator (Nigerian banks buy from us partly because we operate with the same data discipline we sell). Every gateway log row with is_redacted = true flows into a counter on the customer-facing console. NDPC regulatory updates and CBN circulars are watched. A near-miss on the PII redactor triggers automatic policy review.

The Product Improvement Loop reads PostHog funnels, Supabase error patterns, Sentry exceptions, and customer feature requests. Bugs affecting paying customers get founder attention within 24 hours. Minor bugs get agent-drafted fixes with code review. The differentiation audit and NDPA redactor unit tests pass before anything that touches customer data merges to main.

The Strategic Intelligence Loop watches African tech news, competitor changelogs, regulatory developments, partner status changes, and VC ecosystem movement. High-signal events trigger same-day briefings; medium-signal events go in a weekly digest. We are building this so the founder reads one short briefing every morning and stops missing things he should have seen.

These are not aspirational. They are the operating structure from day one. The Compliance Loop already runs every PII redaction through unit tests that were written in week two. The Sales Loop has been recording every BD call since week one. We have the unit tests, the redactor, the audit log, the counter on the customer console, the test sandbox that puts a 4-digit PIN gate in front of any fund movement. The customer-facing console is the Compliance Loop made visible to the buyer.

Why we record everything

Every customer-facing conversation, every internal decision, every material strategic thought is captured. If a conversation was not recorded, it did not happen to Tracent's intelligence layer. The default tooling is self-hosted Whisper running on a Hetzner box that costs less than six euros a month, with a drop folder that any team member can sync into. Recordings get transcribed, transcripts get committed to a private git repository, and an agent processes them into structured summaries the rest of the loops can query.

We do not use Granola. We do not use Fathom. We do not use any third-party meeting bot that sits in calls on our behalf. The reasoning is brand-aligned and policy-aligned at the same time: Tracent sells data sovereignty to its customers, so Tracent operates with data sovereignty in its own internals. A startup that asks Nigerian banks to trust it with BVN-adjacent traffic cannot simultaneously be piping its own BD calls through a US-headquartered analytics vendor.

The same NDPA discipline that applies in production applies internally. Transcripts that mention BVNs, NINs, account numbers, or balances get processed through lib/ndpa/redactor.ts, the same utility the production gateway uses, before any LLM sees them. The DPO retains access to the unredacted originals under an audit trail. Agents only query the redacted mirror.

The blast radius of this choice is small at three people and enormous at fifty. We are locking it in now while the cost of doing so is one conversation per week and a Homebrew install.

Ephemeral software, durable knowledge

A central operating principle: internal software is regeneratable on demand, and we do not pay technical debt on it. Internal CRM dashboards, BD pipeline trackers, partner-tracking spreadsheets, engineering productivity dashboards, investor update generators, all of these get generated when we need them and discarded when needs change. If something would take more than four hours to build internally, the right question is whether the workflow itself is wrong, not whether to invest more engineering time in the tool.

What is durable is different. The customer-facing product is durable: the gateway, the console, the MCP server runtime when it ships, the marketing site. Skill files are durable: every recurring function or knowledge domain has a markdown skill file under version control. The Strategic Compendium is durable. Recordings and transcripts are durable. Customer data in Supabase is durable. Compliance documentation and audit trails are durable.

This split protects against two failure modes that have killed mid-stage African startups before. It protects against AI model improvements making last quarter's tooling obsolete (the model gets better, we regenerate). And it protects against the slow accumulation of internal complexity that, by month thirty, has half the engineering team maintaining tools instead of shipping the product.

No middle management

Tracent has two roles only. Builders, who ship output. And Directly Responsible Individuals, who own a loop. There is no head of operations. There is no engineering manager. There is no VP of growth. If a problem feels like it needs management, the first questions are: can it be a loop, can it be automated, can the existing DRI absorb it.

The named DRIs at the planning stage are the founder, who owns the Sales and the Strategic Intelligence loops; a future CTO with payment-infrastructure background who owns the Product Improvement loop; a future DPO who owns the Compliance loop from month three; and a future GTM lead who owns the Customer Health loop from month twelve. The frontend engineer is a builder, not a DRI. There are no other roles in the first eighteen months.

This is structurally hard to maintain at one hundred people, which is precisely why we are locking it in at three. Adding a manager later is always easier than removing one.

What we're betting

The honest summary is that we are betting on a stack of propositions, any one of which could be wrong.

We are betting that an African deep-tech startup can be built this way, in this environment, with these constraints, and beat the typical African deep-tech fate. We are betting that compliance from day one is a moat, not overhead, because Nigerian banks will buy from the company that already passed the NDPA bar rather than the one that needs eighteen months to get there. We are betting that domain knowledge captured in skill files and version-controlled transcripts compounds faster than headcount, so the company can stay small for longer and still beat larger competitors on operational quality. We are betting that recording everything internally lets us serve customers who care that we operate with the same data discipline we sell to them. We are betting that self-hosted infrastructure on Cassava Cloud and Hetzner, with no paid SaaS in Phase 0 and Phase 1, will produce a tooling cost structure no US-headquartered competitor can match. We are betting that Naira-first pricing, locally hosted gateway, and African languages in the agent interface compound into something Composio and Lunar.dev cannot replicate from outside the continent.

If we are wrong on any one of these, we adjust. If we are wrong on most of them, the bet was bad. We will know which way it went within eighteen months, because compounding either kicks in or it does not.

What could go wrong

A few things we are watching. The Naira could devalue past ₦2,000 to the dollar, in which case our customers' budgets in dollar terms collapse and our hosting cost in Naira terms triples; we have planned conservatively and we have a contingency, but it would hurt. Anthropic or OpenAI could shift their MCP pricing model in a way that makes per-call economics worse for our customers; we are watching their changelogs and we have a path to provider portability. Composio or another global MCP player could decide to seriously enter African payment rails; our bet is that they will not do it with the regulatory and infrastructure investment we are making, but we could be wrong. A new Nigerian regulator could issue a directive that materially changes what gateway operators must do; we read NDPC and CBN circulars within 72 hours of publication for exactly this reason.

We have a small bet on Cassava AI Factory Nigeria launching on schedule, which would let us migrate compute closer to customers and turn our own infrastructure into a marketing asset. If it does not launch, we run on Hetzner indefinitely and the line about sovereign African compute becomes a future tense for longer than we would like.

An invitation

If you are a senior engineer who reads this and finds it interesting, here is how to talk to us. We are not hiring a marketing manager. We are not hiring an engineering manager. We will hire a CTO with payment-infrastructure background in month three to month six, a senior frontend engineer in month three to month four, and a part-time DPO in month three. If any of those is you, send a note to hello@tracenttechnologies.com with a paragraph on the most interesting compliance problem you have shipped, and a link to something concrete you have built.

If you are a CTO at a Nigerian Mid-Tier B2B SaaS company and you have been wondering how to make your platform agent-callable without giving up customer data, the customer console at tracenttechnologies.com is the front door. Sign up, connect an MCP server, watch the redactor at work in the audit log, run a test request, see the human-in-the-loop gate prompt for a PIN before any money would move. If what you see looks like it could serve your customers, the team that takes new integrations seriously is at support@tracenttechnologies.com.

We will publish more posts as the operating loops mature. The next one is about the NDPA Compliance Loop, what it actually does end-to-end, and why we think the gap between policy and implementation is where most African data-protection efforts fail. If that sounds like the right rabbit hole, the blog at tracenttechnologies.com/blog is the right place to come back to.

Tracent operates this way because we believe African deep tech is buildable if it is built deliberately. The five loops, the recording discipline, the ephemeral-software-durable-knowledge split, the absence of middle management for the first eighteen months: these are not preferences. They are the operating thesis we are willing to be measured against, publicly, from week one.