I'm Mohsin Kokab. I've kept regulatory platforms running at 99.99% for DAX40 clients, moved monoliths onto Kubernetes, and led the teams that did it. Today I build with agentic engineering: loop and graph orchestration with verification gates. That is what makes the speed repeatable instead of lucky.
Measured
down 78%
Evidence
Every bar below is a claim with a derivation behind it: what was measured, against which baseline, and where that baseline came from.
down 99%
5× faster
down 91%
down 25%
down 90%
Method
An agent that writes code is not the interesting part. What changes the outcome is where the check sits: put a machine-checkable acceptance condition in front of the generation instead of behind it, and every attempt gets graded without a human in the way. Nothing here makes the generator more accurate. It makes being wrong survivable.
The leverage is upstream: agentic engineering in 2026the six stages around this loop, the published research the gate design rests on, and the randomised trial that argues against all of it
Work
Each one opens as far as it honestly can: some into a write-up, some into the running thing.
Watches every automation you run and turns the night's failures into a morning brief: what broke, why, and the next step for each. Failures are grouped by root cause, and the dead-letter queue replays without double-processing.
A replay checks the item key before it re-executes, so it will not double-send an invoice to make a dashboard green.
An operations console for voice-AI phone agents: every call the agent took, every question it could not answer, and every write-back to the CRM shown as verified by a read-back rather than assumed from a 200.
A write-back counts as verified only once a read-back confirms it. A 200 is a claim; the read-back is the proof.
Six years on a DAX40 regulatory disclosure system, monolith to cloud-native.
Regulated finance: the migration had to land without a disclosure window closing.
LLM orchestration over a job pipeline: parse, score for relevance, draft, track.
The interesting part is not the drafting. It is keeping a hundred concurrent flows recoverable.
Experience
6 roles since 2011: high-availability regulatory platforms for DAX 40 issuers, a cloud migration that cut infrastructure spend 25%, and production voice AI since 2025. The bars are the real date ranges, so the founder-and-lecturer years overlap the employed ones.
14 yr 5 mo
Went independent to build production voice AI and agentic systems end-to-end: architecture, streaming infrastructure, and the tooling around them. Now looking for a team to do this at scale.
Stack
Grouped by domain. The year counts are time spent shipping with the tool in production, not exposure to it. Where a count would mislead, it is left off rather than guessed.
Decision matrix
Every tool here won over an alternative. These are the trade-offs I accepted and the systems they shipped in.
7 of 7 technologies
| Technology | Why & when | Trade-offs | Shipped in |
|---|---|---|---|
| FastAPIBackend | High-perf async Python APIs | Smaller ecosystem than Django | MockGen |
| Spring BootBackend | Enterprise microservices | Heavy startup, verbose config | Regulatory Platform |
| LangChainAI | LLM Orchestration | Fast breaking changes | Job Automation |
| KubernetesCloud | Scale & Multi-tenancy | High complexity | Regulatory Platform |
| ReactFrontend | Interactive UIs | State complexity | Portfolio |
| PostgreSQLDatabase | Reliable relational data | Vertical scaling limits | Regulatory Platform |
| GitLab CIDevOps | Compliance-as-code pipelines | Vendor lock-in | Regulatory Platform |
Patterns
Every pattern here has a cost. These are the conditions under which I think it is worth paying, and the ones under which it is not.
Decomposing monoliths by domain
Async communication via broker
Zero downtime traffic switch
Context injection for LLMs
Legacy -> Proxy -> New ServiceVectorDB -> LLM -> PydanticWriting
Long-form write-ups of problems that took real work to get right, each one argued from primary sources and honest about where the evidence runs out.
AI EngineeringAugust 6, 202612 min read
Benchmarks are saturating and the generator is no longer the variable. What decides whether the output is correct is how little ambiguity survives before code exists, and whether the checks downstream fail independently of each other.
Generative AIJune 3, 20269 min read
Chunking, hybrid search and reranking all matter. But most production RAG systems fail because nobody can say whether a change actually made retrieval better.
Generative AIApril 9, 20268 min read
Constrained decoding gets you schema-valid JSON. It does not get you correct JSON. What broke in two production systems of mine, and what the fixes cost.
Verified LinkedIn recommendations
What a director of engineering, a senior developer and a security leader have written about working with me.
“Mohsin consistently demonstrated exceptional adaptability and technical prowess, quickly mastering complex domain knowledge... his dedication to fostering a collaborative environment was truly commendable.”
“He worked with me in many complex projects from planning phase , architecture, implementation to delivery... he has a deep understanding in Object Oriented Programming, database design and Dockerization.”
“He has a gift for breaking down technical concepts into understandable language... I wholeheartedly recommend him for any leadership role in software engineering.”
Contact
Full-time, in Berlin, across Germany, or remote. Email is the fastest route: I reply within 2 business days.
Twenty minutes, if you’d rather talk than write. Same for a first interview.