MLE | Agentic AI | Founder

MLE Agents & LLMs Data Science Web Math Python Infrastructure Claude Gemini

🛡️ LogGuard – CI logs consume LLM context & input; lg run compresses failure-preserving.
🔬 Indago – Agentic investigation packs for coding agents over frozen ML incident evidence. Practitioner outreach: complex prod-ML incident RCA is rarer / less painful than assumed.
📖 Excerpta – Nonfiction fluff + floating AI summaries break trust; paragraph-anchored key points, local only.
🤖 NotionQuery – Notion API filters are tedious to build; natural language → validated Python queries.
📬 NotionAgent – Captured notes rarely become actionable; messy inbox → prioritized Notion tasks.
📦 FashionMNIST-Ops – Production-style ML pipeline: Docker, tests, profiling, reproducible training.
⚖️ Uni-Judge Research – Can LLM judges replace classical classifiers with only dozens of labels?

🔀 RouterExperiments – When is an expensive LLM actually necessary? Cost-aware router benchmarks.
🧪 Synthetic Routing – Synthetic data is cheap — does it improve routing or teach shortcuts?
🎯 DPO-RAG-Opt – Small LLMs break fixed JSON schemas; measures whether DPO fixes structured output under RAG.
🔬 LogGuard Experiments – R&D monorepo behind the LogGuard v1.0.0 release.
🎧 manga-tts – English manga audio overlay: vision extract → TTS → player (personal research dogfood).
📡 personal-tracker – Reconstructable text-of-work (ActivityWatch + keystrokes) for LLM-pasteable daily stories.

🧪 production-ml-lab – Learning artifact: compressed production-ML immersion lab (Indago substrate — not a product).
💬 Messenger – Real-time messaging with auth, persistent storage, and WebSockets.
⛈️ Bad-Weather-Notifier – Location-specific severe-weather alerts.
🎮 Geometry-Dash-AI – RL environment for autonomous Geometry Dash gameplay. (Paused.)
🗺️ Visual Long Planner – Archived. Visual+chat plan graphs; falsified — planning wasn’t the bottleneck (tasks weren’t that complex).
🔬 Skepsis – Archived. Experiment dashboards hallucinate metrics; agent harnesses (Cursor) made a custom investigator redundant.
🃏 Anki-Accelerator – Archived. Card intent is too personal; writing the card is half the learning.
🧠 Scientific Reasoner / Avatar / Trace Analysis – Archived. Memory OS, judgment-avatar, and mechanical-trace clustering — all falsified as products (WTP / Cursor substitute / weak market pull).

🧠 LeetCode – Top 5%
📊 Kaggle – Multiple top 5-10% playground appearances

Building agentic systems, experimenting with LLMs and architectures, and solving all friction around. Philosophy: Build. Grow. Pursue.

• Open To: startup work, co-founding, research collaborations, and ML Engineering / Agentic AI roles.
• Connect on Twitter/X • Telegram