Mohamed Diomande — New York
Research, built
into systems.
I do machine learning research on speech, writing systems, and autonomous agents, and I build the infrastructure that puts the research to work. The papers live next to the systems they describe.
Capabilities
What I arrive with
Machine learning research
Speech recognition for low-resource languages. Writing systems as model infrastructure: tokenization, script visibility, phonemically interpretable evaluation. Transfer learning and CTC decoders trained under real data constraints. Reward modeling over agent trajectories.
Agent systems engineering
Multi-machine orchestration: a personal mesh of Macs, a Windows GPU node, and cloud workers coordinated through deterministic supervisors, message streams, and cron-governed loops. Trajectory scoring, typed skill libraries, persistent memory systems, autonomous multi-hour build sessions.
Real-time perception and graphics
Pose estimation pipelines from commodity cameras and depth sensors. Audio-reactive and body-reactive generative visuals in Unity and TouchDesigner. Multi-iPhone camera meshes with on-device capture orchestration. Gaussian splatting on consumer hardware.
Product engineering
iOS with SwiftUI, shipped through TestFlight. On-device inference: whisper.cpp, quantization, Apple Neural Engine serving. Web with Next.js. The discipline of carrying a model from training run to a phone in someone's hand.
Operating physical businesses
Specialty coffee service in New York under Buf Barista. Plant-based milk distribution under Koatji. Field-tested in inventory, events, sales routes, and the unforgiving feedback loop of selling something real.
Research
4 programs
Language as Infrastructure
N'Ko is not a decorative rendering of Manding language. For machine learning systems it is computational infrastructure: it determines what tokenizers can represent, how acoustic evidence aligns to symbols, and whether a reported error rate measures speech recognition or merely agreement with an inherited orthographic convention.
12 entries
Agents That Account for Themselves
Autonomous systems earn trust through verifiable provenance, measured trajectories, and typed composition, not through claims. This program builds the accounting layer for agent work: what was done, what evidence supports it, and what it was worth.
7 entries
Embodied Trajectory Systems
Motion, conversation, graph traversal, and personal trajectory are treated as paths through state space. The common object is not the sensor, the text, or the model output; it is the trajectory, measured by geometric signals that can be replayed, scored, and governed.
5 entries
The Absorption Loop
Staying state-of-the-art is a process, not an event. Every day, curated papers are triaged against an explicit architecture lens and judged with one of five falsifiable verdicts: absorb the technique, test against our build, claim rivalry with evidence, watch with an explicit trigger, or skip. The log below is public because research practice should be inspectable.
1 entry