Lou Diamond Morados

AI/ML Engineer · Eden (Ed3n Ventures)

Apr 2026 – Present

• Own the full lifecycle of proprietary Small Language Models — synthetic data generation, PEFT fine-tuning (QLoRA, DoRA), DPO alignment, evaluation, and release — selecting and adapting base models per business use case and taking each from experiment to shipped artefact on a repeatable path.
• Rebuilt the fine-tuning data pipeline around single-source regeneration from one teacher model, so training runs became auditable and repeatable rather than one-off.
• Benchmarked offline sequence-level against on-policy logit-level distillation (TRL DistillationTrainer) and specified an open-weight 27B teacher path that held task accuracy at materially lower inference cost.
• Lead the build of an AI-native mobile application — React Native client on Go services — including an LLM gateway and smart router that fails over to an alternate model when a provider's usage limit is reached, keeping the product available without changing calling code.

Machine Learning Engineer · WizyVision

Oct 2025 – Mar 2026 · intern, then contract

• Engineered on-device generative AI pipelines for rugged industrial hardware, owning the flow from raw input through preprocessing to inference.
• Benchmarked industrial against consumer mobile devices to establish the latency and reliability baselines deployment decisions were made against.
• Designed a state-based tool-guarding architecture constraining what the model was permitted to do at each step, enforcing data integrity across multi-step reasoning — on-device inference on field hardware, where a wrong tool call is not quietly recoverable.

Software Engineer Intern · Sahei Core Technologies

Jun 2025 – Jul 2025

• Built a credit-scoring model for a cooperative lending platform — sourced the approach from the microfinance literature, integrated it into the .NET risk-assessment flow, and presented and defended the methodology to the General Manager and to client cooperatives.
• Designed and owned the PostgreSQL schema — tables, constraints, audit triggers — and refactored the query layer to parameterised, injection-safe statements.

nexum

Real-time trend intelligence off the Bluesky firehose, on a $10 box.

Consumes the Bluesky Jetstream firehose with asyncio and websockets, produces to Redpanda with backpressure and time-cursor resume, and detects hashtag bursts in a Quix Streams dataflow — hopping-window counts over RocksDB-backed state, z-scored against a rolling EMA baseline. Posts archive to date-partitioned Parquet as the future training corpus; a FastAPI/SSE dashboard shows live throughput and the burst feed. The point is to replace the cloud API call in the hot path with a self-distilled ~270M classifier at measured teacher agreement and a measured p99.

Python · Redpanda · Quix Streams · RocksDB · Parquet · DuckDB/dbt · FastAPI

GitHub

PowerWatch

2M+ household energy readings turned into a tested dimensional model.

End-to-end ELT warehouse on Google Cloud: raw consumption data loaded into BigQuery, transformed by dbt into a galaxy schema with automated data-quality tests that fail the pipeline on anomalous input, surfaced in a live Looker Studio dashboard, and queried in plain English through a Vertex AI natural-language-to-SQL assistant constrained to the warehouse schema.

BigQuery · dbt · Looker Studio · Vertex AI · Python

GitHub · Live dashboard

CreatorPulse

A daily service that runs with nobody logged into the box.

A systemd timer on a rented Linux VPS fires a Python job that pulls public creator statistics, writes them with full daily history into SQLite, syncs a live Google Sheet, and posts an 08:15 digest to Discord that doubles as a heartbeat — flagging ±20% movement and that run's failures, and answering /creator and /status in-channel.

Python · Linux VPS · systemd · SQLite · Google Sheets API · Discord

GitHub

Owlark Honorable Mention, Developer Camp Manila 2026 — 1 of 8 teams awarded from 53 projects

An offline AI tutor for the cheapest Android phone in the room.

Offline-first educational tutor for low-end Android devices in areas with no reliable connectivity. Hot-swaps 10MB subject-specific LoRA adapters into a shared Gemma 3 base, serving several specialised models from one set of weights inside a constrained memory budget rather than shipping a model per subject.

Gemma 3 · LoRA · Edge AI · Android

Results

VoltGuard 3rd place, Meralco IDOL 2025

Ask your appliance logs a question in plain English.

Agentic pipeline bridging a locally-hosted Llama 3.2 model to PostgreSQL over Model Context Protocol, exposing querying as a bounded tool set so appliance logs can be interrogated in real time without handing the model free rein over the database.

Llama 3.2 · Model Context Protocol · PostgreSQL · Python

Coverage