Kinetix: An Agent Harness Inside a Clinical System
Sixty tools, twenty-nine of them mutating, inside a system holding injury records. The Go that enforces one rule: the agent inherits the caller.
Biomechanics from Video08
I build the layer between a language model and software that already exists.
Five-plus years on the same stubborn question: how do you take something that works in a notebook and make it survive real users? Mostly that has meant backend systems, production machine learning, and the harness layer in between. Lately a fourth thread has crept in, machine learning for biology, where the datasets are small, the labels are noisy, and being honest about your numbers matters more than beating a leaderboard.
Now
Since Aug 2026
Senior Software Engineer, Mevrik
An AI customer-experience platform used by telecom operators and several hundred smaller businesses. I work on the AI side: agents, retrieval, and the serving infrastructure underneath.
Since 2024
Founder, AlgolyzerLab
A small studio for applied AI that does not fit a product roadmap: sports science for elite cricket, clinical software for doctors, agriculture AI, and research engineering.
What I think about
01
Agent harnesses
A model is only as useful as the surface you hand it. Most teams give an LLM raw tables and generated specs instead of the intent-named operations their own code already has.
02
Machine learning for biology
Anti-inflammatory peptide prediction, and structure-aware virtual screening of natural products against understudied viral proteases. Both taught the same lesson: the dataset you choose decides the result long before the model does.
03
Recent writing
An Agent Is Data, Not Code
Store agents as versioned, validated rows instead of writing them as classes, and rollback, diffs, audit, teams and lessons come free. Here is the bill.
Agent Harness04
What the System Refuses to Measure
Three measured refusals in a markerless movement pipeline, the gaps that are still open, and why a clinician should prefer software that says not measured.
Biomechanics from Video07
Finding Sick Plants From the Sky: DINOv2, a Linear Probe, and a Farmer With a Drone
A frozen vision backbone, a linear probe trained on forty photos, and a three-tier fallback so that a drone survey works on day one.
Vision in the Real World03
Triangulation, and How Many Cameras You Actually Need
The multi-view path: DLT in enough detail to implement, how many cameras earn their place, and three lessons about evaluating a solver whose truth you cannot see.
Biomechanics from Video06
From 26 Keypoints to Clinical Metrics
Two calibrated cameras, DLT triangulation, trajectory filtering, and honest uncertainty intervals: what turns a pose demo into a record a clinician will sign.
Vision in the Real World02