Fredrick Otieno

Software & AI engineer. Introvert who ships. Building in public.

[Fredrick Otieno @DataAnalyt2539]

AI agents can think. They still can't pay.

So I built Loki: an MCP server that lets your agents send M-Pesa payments and use virtual cards, inside the caps, budgets and allowlists you set.

Built on @SafaricomPLC Daraja. Sandbox-tested. Waitlist below.

#MPesa #AIagents #MCP

[Fredrick Otieno @DataAnalyt2539]

Agents will retry. Networks fail, models loop. If a payment tool isn't idempotent, a retry becomes a double charge. With Loki (sandbox-tested MCP payment gateway) every call carries a key, and per-payment caps plus monthly budgets return hard denials, not warnings. #MCP

[Fredrick Otieno @DataAnalyt2539]

@YogeshBaks53995 Per-request keys are right, but also persist the key with the result, not just a seen-set, so a retry after a timeout returns the original outcome. For reconciliation, treat Stripe webhooks as source of truth and keep the stream's ledger entries pending until confirmed.

[Fredrick Otieno @DataAnalyt2539]

I'm less interested in which model wins the leaderboard.

I'm interested in where judgment still lives when the model is in the loop.

If your system can't say "I don't know," it will confidently ship the wrong answer.

[Fredrick Otieno @DataAnalyt2539]

@rY3kop The "fix what you broke while hardening" loop is real. Detection software that survives its own false positives is the boring half of shipping security tools.

[Fredrick Otieno @DataAnalyt2539]

@ClementDelangue Open weights that run well on Mac Studios matter more than the launch headline. Local-capable models change what you can ship without a cloud round-trip on every call.

[Fredrick Otieno @DataAnalyt2539]

@ThePrimeagen Dotfiles as agent setup docs is underrated. The useful part is encoding the non-obvious machine preferences once so every new session starts from the same baseline instead of re-explaining it.

[Fredrick Otieno @DataAnalyt2539]

@WOkparaji74619 Logging detection rate next to accuracy will make the dim-light and blur runs much clearer. Looking forward to the matrices.

[Fredrick Otieno @DataAnalyt2539]

@sparrowbrainz Solid stack — SELECT-only user plus a single-statement parse. That puts the hard limit in the DB and the parser, not in a prompt the model can talk around.

[Fredrick Otieno @DataAnalyt2539]

@__ciox__ @engimeets That double-tax gap is the real cost of the Delaware path from Kenya. Worth modeling payouts and KRA exposure early, before the first US invoice, so the entity choice still pencils once both sides collect.

[Fredrick Otieno @DataAnalyt2539]

@CharlesHNR_ @make_hq Nice flow. I would add a fallback branch so leads still get the instant alert if the OpenAI step errors or times out. Speed-to-lead breaks quietly when one hop in the chain fails.

[Fredrick Otieno @DataAnalyt2539]

@levelsio @Microsoft @XBOX @satyanadella @asha_shar Fan ports made out of enthusiasm are free marketing for the original games. Embracing them keeps the community building on the brand, and that goodwill is worth far more than the takedowns would protect.

[Fredrick Otieno @DataAnalyt2539]

@levelsio The uncomfortable part is that customers never agreed to their workflow data being resold. Once that is a known exit path, data-processing terms and deletion rights become a real buying criterion for SaaS.

[Fredrick Otieno @DataAnalyt2539]

@WOkparaji74619 Good result on dim light. Since scores only count frames where the hand was detected, I would also log the detection rate per condition next to accuracy. That shows whether lighting hurts the detector or the classifier.

[Fredrick Otieno @DataAnalyt2539]

Some skills are just worth giving a model

[Fredrick Otieno @DataAnalyt2539]

@ArifenAziz Pixel-art portfolio that actually ranked is a strong signal. Interactive plus sound is rare for personal sites — nice craft.

[Fredrick Otieno @DataAnalyt2539]

@agentic_bala Yes. Graceful failure with a clear skip reason is the difference between a demo and something you'd trust with money or tickets.

[Fredrick Otieno @DataAnalyt2539]

@rileybrown Voice-first while walking is a real workflow shift. The hard part is keeping the agent scoped so a yap doesn't spawn five half-finished sessions.

[Fredrick Otieno @DataAnalyt2539]

@theo Internal token spend as a product feedback loop is underrated. If your own agents can't burn that much usefully, the product surface probably isn't agent-shaped yet.