Bhekani Khumalo
Senior Software Engineer · AI and distributed systems
Summary
I'm a senior engineer at Contentful. In 2025 I worked on the semantic layer behind the platform's AI features: pipelines that vectorise millions of customer content items, the vector store on Aurora, and the APIs that search, recommendations and duplicate detection call. The ML engineers proved the approach; I made it run in production, which meant batching, fallbacks, cost ceilings, and deciding what happens when a model call times out behind a synchronous customer endpoint. Since January 2026 I've been on Workflows, leading the work to extend workflows and automations to new entity types across six services.
Outside work I've shipped blah.chat (multi-model chat with RAG), easydeck.app (AI slide generator) and mxr.sh (an agent-first email client in Rust). They have users. At Notto, a credit scoring company, I'm the part-time CPO. There's no product engineering team yet, so I also wrote their portfolio analytics product and loan-book agent, both now in customers' hands.
Ten years in software, seven building SaaS. Before software I was a civil engineer in Zimbabwe designing water networks, which is probably why I still start by asking where a system fails first.
Skills
- Languages
- Python, TypeScript, JavaScript, Go, Rust, SQL
- AI/ML
- LLM integration and orchestration, agentic workflows, RAG and retrieval, embeddings and semantic search, vector databases, multi-model and multi-modal systems, voice AI, evaluation and benchmarking
- Backend and data
- Distributed systems, microservices, event-driven architecture, batch and streaming pipelines, relational databases, caching, warehouse-backed reporting, REST and GraphQL APIs
- Infrastructure
- AWS, infrastructure as code, Docker and Kubernetes, CI/CD, observability and incident response
- Frontend
- React, Next.js, Svelte, TypeScript
Experience
Contentful
Composable content platform, London. Senior Software Engineer across three teams.
Workflows
- Lead engineer for extending Workflows to the new entity types in Contentful's experience platform. Wrote the technical design across six services, refactored the Workflow Management API runtime around subject adapters, and shipped the storage migration with an idempotent, dry-run-first backfill.
- Shipped Automation Builder features: entry query filters through the execution pipeline, feature flags for new action types, and failure handling that keeps error context for support.
- Routine production debugging across the stack. One recent case: a 403 storm traced through three services to environment-alias handling in the web app.
- Interview engineers for the team, usually leading the system design round. Wrote the org-wide conventions for AGENTS.md and skills, and the team's golden-context docs for coding agents.
AI Platform
- Led the design of the semantic service behind AI features across the platform, including the vector storage layer on Aurora, with index tuning and partitioning so retrieval latency stayed flat as the corpus grew.
- Worked alongside ML engineers to take model work from prototype to production: batching, validation, cost controls, fallbacks and the constraints of serving models synchronously behind a customer-facing API. Set up the monitoring and alerting the team now runs on.
- Built Kafka, Lambda and Step Functions pipelines that keep embeddings current across millions of content items in near real time.
- Shipped the retrieval endpoints for semantic search, recommendations, duplicate detection and contextual references.
Business Foundations
- Led a cross-functional initiative to build organisation-level audit logging, which meant getting services across the platform onto one event contract to meet enterprise security and compliance requirements.
- Consolidated billing and usage reporting into a single warehouse-backed platform, replacing brittle pipelines and removing manual work finance and support repeated every month.
- Fixed a bottleneck in a high-traffic gateway service, taking query times from seconds to milliseconds.
Notto Africa
Alternative credit scoring for African lenders. Fractional CPO, part-time alongside Contentful. No product team yet beyond me; I set direction and write the code.
Fractional CPO
- Built Notto Analytics and Notto Agent, both in production with customers on a live book of close to a million loans.
- Analytics ingests a lender's disbursements and repayments daily and computes vintage PAR, roll rates and yield, with drill-down by product, cohort and segment.
- Agent answers open-ended questions about the book, sends scheduled insights, and runs multi-step analysis workflows. It works on any loan book, not only ones Notto scores.
- It operates in regulated lending, so answers are verifiable rather than trusted: it plans the analysis, runs SQL against a read-only copy of the data, checks the result against the evidence, and keeps plan, query and trace for audit.
- Before customer testing, added a spend ledger, per-tenant and per-user limits, rate caps and thread compaction.
- Wrote the daily ingest underneath: lender CSVs arrive over SFTP, load into TimescaleDB, feed the dashboard. Runs on-prem inside the partner's own data centre with no outbound network.
- Own the product definition across the four products. Engineering, credit science, commercial and marketing now describe the product the same way, which they didn't in April.
Anaplan
Software Engineer, Integrations. London.
Software Engineer
- Led a ground-up rebuild of the DocuSign extension: migrated Flow to TypeScript, rewrote the test suite with React Testing Library and Cypress, and moved state management to React Context.
- Rebuilt a legacy API as a GraphQL server on Apollo, so the team could change it safely.
Sigma Digital
Senior Software Developer and Team Lead. Remote, South Africa.
Senior Software Developer / Team Lead
- Primary technical leader for a distributed team building custom software for enterprise clients, owning scoping, architecture, delivery and the client relationship.
- Set the code review, testing and release practices the team then ran on across timezones.
Earlier: Frontend Software Engineer, The Spry Group, Sep 2019 to Jan 2021 (concurrent consulting; React, Gatsby, TypeScript). Before software: GIS Developer at Esri South Africa and RAMM Technologies, then Civil Engineer at City of Bulawayo and Kuma Consulting, designing water networks and managing a $1.8M sewer rehabilitation project.
Independent products
50+ models behind one interface, switchable mid-conversation, bring your own key or self-host. RAG over uploaded documents, memory, multi-modal input, agentic workflows. An orchestration layer routes each task on cost, capability and latency, with per-request cost shown to the user. Next.js, Convex, vector search.
Describe a deck, get finished slides with original generated graphics as PowerPoint or PDF. Research, structuring, visuals and layout are each a discrete step with its own validation, so one bad step doesn't poison the whole deck.
A background daemon syncs Gmail, Outlook, Microsoft 365 and IMAP into SQLite, keeps a Tantivy search index, and exposes everything over a JSON-over-Unix-socket API. The terminal UI, CLI, web app, MCP server and agent skill are thin clients on that API. Every mutating command has a dry run. A deterministic pre-send pass checks for missing attachments, leaked secrets, reply-all mistakes, unknown recipients and tone drift. Search is BM25 plus local embeddings, all on-device.
Also: worthyourtime.xyz (film recommendations from a learned taste model that explains its verdict), faithbench.com (public LLM benchmark with a multi-stage LLM-as-judge evaluation pipeline and a held-out test split), interviewoptimiser.com and referenceoptimiser.com (real-time voice AI for mock interviews and reference checks, with Hume AI), dealbase.africa (AI extraction of African startup funding data), imali.cc.
Speaking, mentoring, education
- Talks
- Svelte Society London, 2022, "How I built pro-search.io with SvelteKit". DevConf South Africa, 2020, "What civil engineering taught me about being a good software developer".
- Mentoring
- Coding Mentor, Andela Learning Community (Google scholarship initiative). External Academic Supervisor, National University of Science and Technology.
- Education
- BEng (Hons), Civil and Water Engineering, NUST, Bulawayo, Zimbabwe, 2009 to 2014.