By Geetansh Jindal.
I build production AI systems and write down what breaks after they ship. A decade of infrastructure work sits underneath most of it.
I build systems that turn messy human input — speech, documents, argument — into something a machine can act on. Then I spend most of my time on the part that comes after: what happens when the inputs get noisy, ambiguous, or adversarial and the thing still has to work.
Two of those are companies.
CallOptix. Co-founder and CEO. Sales and support teams record everything and learn almost nothing from it. The objections reps fumble, the moment a deal goes cold, the patterns the top closers use without knowing they use them — all of it sits in audio nobody has time to review. CallOptix listens to every call in 100+ languages and turns it into structured signal: transcription, diarisation, agent scoring, compliance monitoring, pipeline. It runs for call centres, BPOs, and sales teams where the phone is still the primary channel.
Jury Analyst. CTO. Legal tech for US trial teams — jury simulation, damage range prediction, and the research that goes into voir dire. Technology strategy, product, and engineering.
Before either: a decade of infrastructure. Server administration, then DevOps. CI/CD pipelines, fifty-odd production servers across AWS and bare metal, monitoring, security hardening, incident response. Certified Ethical Hacker, and enough penetration testing to have opinions about how systems actually fail. That decade is why I'm hard to impress with a demo. I spent it on the side of the wall where things page you at 3am.
What I'm interested in
The interesting problems in AI are not in the model. They're in the six inches around it.
A retrieval system that returns the right document 94% of the time and is confidently wrong on the other 6%. A transcript that holds up until the speaker switches language mid-sentence. A date parsed out of a scanned report that silently reorders a timeline and nobody notices until an answer comes back wrong. An agent that behaves in evaluation and does something else entirely on a Tuesday afternoon under real traffic.
Latency, ambiguity, edge cases, human behaviour, and the invisible system decisions that decide whether a model produces value or noise. AI becomes meaningful when it improves a real decision. Not when it looks impressive in isolation.
What's on this site
Projects, mostly open source. Notes from systems that are actually running. Occasionally an investigation that got long enough to need its own page.
The bias is toward self-hosted, Postgres-shaped, and boring where boring is correct. If a write-up here says something worked, it worked somewhere real. If it says something broke, I broke it.