@isHeSatoshi

Summary

LLM research engineer working on agents, evaluation harnesses, and inference. Improving pretrained autonomous-driving decision models with reinforcement learning, shipping production RAG systems, and distilling models down to run on consumer hardware. Most interested in how attention and quantization actually behave at runtime, and where the architecture breaks.

Education

Experience

• Improved pretrained autonomous-driving models by building reinforcement learning environments and simulation workflows to investigate decision-making failures.
• Designed RL environments that recreate those conditions and trained policies through trial and error on H200 nodes to fix the identified decision failures.

• Built a wearable-sensor model that flags when a person's body switches from aerobic to anaerobic breathing during exercise — useful for athletes and fitness researchers.
• Designed hand-crafted time-series features (heart-rate variability, sweat-rate trends, motion intensity) to predict muscle cramps minutes before they happen.
• Rebuilt the data pipeline with an automated validation harness so biomedical researchers could iterate on the model without manual preprocessing each round.
• Wrote plain-English reports and demo notebooks so clinicians and product managers could interpret model behavior without reading the code.

• Replaced a chunk of invideo's human-handled support tickets with an AI chatbot built on OpenAI and LLaMA, grounded on internal help-center articles via a vector search index.
• Shipped the full Flask + React chatbot with a thumbs-up/down feedback loop that automatically tunes prompts weekly based on real interactions.

Projects

• A deterministic state machine goes in front of any object detector: idle every 0.4 s, fire on straddle, hot-cap every 0.1 s while a task entity is crossing the band midline.
• Recall is bounded by Lemmas 1–3. Any midline crossing the detector would have seen at full rate gets reported with timestamp error ≤ 0.4 s + 0.5/f_ps. Property tests cover the boundary cases.
• World memory holds identity and motion across the idle gaps, so nothing upstream has to track anything.
• On 8 real aerial clips (YOLOv8s @ 640, VisDrone + VIRAT, RTX 4060 Ti): 25–42% of windows sampled, up to 78.7% less wall-clock. The GT eval covers every midline crossing at a 0.3 s cadence, 60 of 60.
• Training-free, model-agnostic, tracker-free. A patent and literature search turned up no prior art on that combination.

• Built a Gymnasium-style training harness where an agent learns new task families from real execution feedback alone — no gold labels in training and none in the safety gate.
• Runs agent actions inside an isolated sandbox with AST safety checks, timeouts, and error capture, so reward comes from code actually executing rather than a judge model.
• Added a holdout-veto safety gate that rolls back any LoRA adapter update hurting held-out accuracy, and replay-based memory to control catastrophic forgetting across task streams.
• Evaluated on MBPP and HumanEval with deterministic content-hash train/holdout splits; the main finding was that swapping the LoRA adapter each step beats self-reflection that sees the gold answer.

• Designed a custom 4-mic XMOS audio PCB so the assistant can hear you from across a noisy room (beamforming plus on-chip noise suppression), and wired a distributed ESP32 plus Raspberry Pi network for low-latency audio, wake-word detection, and command routing across rooms.
• Benchmarked 3 versions of a 2B-parameter LLM (off-the-shelf, fine-tuned, and custom pre-trained) and shipped an end-to-end pipeline (mic, DSP, wake router, LLM intent, action) with a live web dashboard for debugging in real home environments.

• Built SmallJev, an open-source 2.5B-parameter variant of TypeSafe's Jev, designed to bring Jev-like intelligence to consumer hardware. Designed a span-pooled semantic scoring architecture with one forward pass and zero generated tokens, targeting devices from an RTX 4060 to Raspberry Pi 5 and phones. Ranked #7 on JevBench with 68.5/100, competing against models up to 35B parameters.

• Solved the problem of businesses struggling to navigate huge BIS standards documents by building a natural-language RAG system that maps product descriptions and questions to the relevant IS codes. Achieved 98% Hit Rate@3, 86% Hit Rate@1, 0.939 MRR@5, and ~1.1s average latency on a 100-query evaluation set using hybrid dense + BM25 retrieval with deterministic ranking. Production system is now used by BIS and tens of thousands of users; built and open-sourced the public release.

• The agent works from symptoms only: tools return data, never the chaos scenario name, and apply_production_config / rollback_production need explicit approval before they touch a real deployment.
• Per-GPU-node controller wraps vLLM with metrics, logs, replay, and eval behind an MCP server. SLOs and evidence rules are pinned in policy.yaml so the agent cannot talk itself past them.

• Reduced deployment from 34 GB to 16 GB and generated 512 px images in 1.45 s through DMD-based Qwen-Image-2.1 distillation, reaching Nano Banana class image generations.

Skills

Community

Hackathons