Work

Applied NLP & Data Pipelines / In production — Prospecto (current role), no public repo

Government tender data is public but not usable: it's spread across 20,000+ agency websites with no shared schema, no shared identifiers for who's bidding on what, and no reliable way to tell a reissued tender from a genuinely new one. This is the pipeline I built at Prospecto to fix that: distributed crawlers for discovery, NLP/LLM-assisted extraction to turn unstructured notices into structured records, and entity resolution to unify the same supplier or buyer across sources that spell its name three different ways. The output feeds procurement teams a monitored, deduplicated view of tenders instead of a pile of PDFs.

Applied RAG & Reasoning / Public on GitHub

Most legal-tech LLMs act as black-box oracles. Compliance Agent flips the architecture: it uses LangChain and ChromaDB for regulatory retrieval, but screens policies using a lightweight DDL-style reasoner to evaluate explainable YAML rules. It acts as a deterministic screening layer that surfaces risk indicators and policy drift for human experts, rather than attempting to give unsupervised legal advice.

Statistical NLP / Public on GitHub & PyPI

Redlines tell you what words changed; LegalDrift tells you if the substantive meaning changed. By leveraging statistical NLP and embedding-based monitoring, this tool detects semantic drift across legal document revisions. It allows contract managers to verify that renegotiated agreements haven't silently introduced new obligations, treating legal documents with the same observability we apply to software state.

Database Engineering & ASTs / Public on GitHub

String-based diffs for database schemas are notoriously fragile. DiffDDL solves this by parsing Data Definition Language (DDL) statements into an Abstract Syntax Tree (AST) and computing the minimal operational graph to transition between states. This provides a deterministic foundation for LLM-driven database agents, allowing them to propose and validate schema migrations safely without dropping state.

LLM Fine-tuning Research / Experimental — design + math implemented, no public repo yet

DoRA decomposes pretrained weights into magnitude and direction, then applies LoRA to the directional component. This project replaces the linear low-rank adapter with a KAN, using learnable B-spline activations instead of fixed linear projections. The result is a PEFT method that captures nonlinear relationships in the weight residual space, adapting patterns that vanilla LoRA systematically misses.

Agentic Systems & RAG / Prototype — exercised on a synthetic corpus, no public repo yet

Standard RAG retrieves semantically similar chunks. Graph RAG traverses entity relationships. Agentic RAG iterates and self-corrects. This project combines all three: an MCP server exposes the knowledge graph, a planning agent decides whether to retrieve by similarity or traverse by relationship, and a critic agent evaluates faithfulness before returning. The synthesis handles multi-hop questions that defeat flat vector search.

Inference Optimization / Prototype — orchestration loop implemented, benchmark numbers not yet public

Speculative decoding uses a small draft model to propose tokens that a large model verifies in parallel. This project lifts that idea to the reasoning level: a lightweight draft agent proposes the next N reasoning steps, and a larger verifier agent accepts or rejects each in a single forward pass. The target is a wall-clock speedup on long multi-step traces with no accuracy loss, the same way token-level speculative decoding works; that target is not yet backed by a public benchmark.

Agentic Systems & Reliability / Design only — not implemented, no public repo

A 5-agent pipeline of 90%-accurate agents produces 59% end-to-end accuracy: the math compounds against you, not for you. Most of that loss isn't model capability, it's unchecked error propagation between steps. This is a design for a deterministic gate placed between every agent handoff: validate the output against a schema, reinject the original task state so context doesn't drift, and route anything below a confidence threshold to a human queue instead of letting it flow downstream silently.

LLMOps & Evaluation / Design only — not implemented, no public repo

An eval test case is a compressed record of a moment the system failed, which means a deploy pipeline that doesn't re-run that set on every change is shipping blind. This is a design for a CI gate: run the production-failure-derived eval set (RAGAS faithfulness, retrieval hit-rate, schema compliance) before merge, block on regression, and pipe every new production failure caught by tracing straight back into the eval set as a new case.