YuktYantra builds and benchmarks machine learning systems where real physical conditions matter: autonomous computer vision for vineyards, high-concurrency speech-to-speech inference, and robot manipulation datasets.
Each programme starts with a practical problem, tests against physical data, and publishes its methods and failure modes.
Season-long smartphone video imaging paired with hand measurements across a working Thompson Seedless vineyard in Sindagi taluk. Building a spatial image archive to evaluate canopy sizing, disease detection, and Brix sugar development.
Researching full-duplex speech-to-speech models that listen and speak concurrently. Our target: serve multiple concurrent phone calls per GPU under an 80ms frame deadline, with acoustic adaptation for 8kHz narrowband telephony and Indic languages.
Sim-to-real procedural vine pruning models paired with a mobile "Skill Challenge" capture concept. Using on-device camera raters as reward models to score physical manipulation attempts while enforcing on-device DPDP consent.
Models that cannot run within strict latency, memory, and energy limits do not reach a farm or an edge device. We treat hardware execution as a first-class design constraint.
Eliminating kernel launch gaps through replayable CUDA graphs, bounded sliding-window KV caches with attention sinks, and static tensor shapes.
Procedural vine simulators with rule-based agronomic rewards, generating diverse training variations before evaluation on real vineyard photos.
Quantization to INT8 and FP8, pruned architectures, and speculative frame drafting tested for latency inside real robot control loops.
YuktYantra is led by Arun Patil, a systems ML engineer with over 17 years of experience spanning frontier LLM inference, specialized FPGA neural network kernels, enterprise-scale retail computer vision across thousands of stores, and embedded camera imaging.
Intellectual property includes 11 patent filings (8 granted) and published machine learning research in transformer architectures.