Atlas Josepheeeee

👋 Hi, I’m Atlas

AI/ML Systems Engineer · Agent & LLM

I build machine-learning systems that connect models with real tools, data contracts, evaluation loops, and reproducible engineering workflows.

What I work on

Agent systems LLM post-training NL2SQL
Document AI Data pipelines Evaluation

What I care about

Clear interfaces · deterministic guards
Leakage-safe evaluation · honest conclusions

✨ Selected work

🛒 Shopping Agent Post-training

LoRA SFT + veRL GRPO for long-horizon shopping agents that must search, verify, select variants, and purchase.

Highlight: Baseline 1.0% → SFT 57.0%; GRPO 58.5%, with the current SFT→GRPO gain not statistically significant.

🧩 DeepVision NL2SQL

Stateful natural-language querying with table routing, schema-aware semantic parsing, constrained DSL, deterministic SQL compilation, and read-only MCP tools.

Highlight: LLM handles semantic mapping; code owns SQL structure and safety boundaries.

📄 AFAC Document Repair

Conservative financial-table structure repair using directional morphology, line detection, chunk-aware processing, and regression checks.

Highlight: Team result: B-rank #7; my focus was table-header repair and regression coverage.

📊 Seagate Fail Monitor

A local, synthetic-data reproduction of a manufacturing quality-monitoring workflow from source tables to dashboard, alert outbox, and traceable results.

Highlight: Runnable backend + frontend demo with deterministic failure injection and recovery paths.

🧰 Toolbox

🧭 How I build

question → explicit data contract → small reproducible run
        → deterministic checks → paired evaluation → honest conclusion

I prefer a useful limitation over an impressive but unsupported claim. Each project README includes the current implementation boundary and the fastest path for a technical reader to reproduce or inspect it.

_{Thanks for stopping by · Explore the repositories or start a conversation on GitHub.}

Navigation Menu