Open to Forward Deployed & Lead AI rolesCV

Forward-deployed AI engineer

I build AI systems that survive production.

I embed with enterprise teams — insurers, BPOs, media, education and the public sector — turn a vague AI ambition into a scoped plan, and ship governed RAG, agent and LLMOps systems into their production.

Download résumé See the work

Now: Intelligent AI Solutions (UK)Ex-AIA Digital+Ex-TDCX AI

6+ yrs · 30+ production projects
  • 95%less FTE for Next-Best-Action model developmentAIA Digital+ · peak reduction
  • 10–30×document-processing throughputDocument Intelligence Pipeline
  • ~40%retrieval cost reductionHybrid RAG
  • 20%+faster agent time-to-proficiencyTDCX AI
  • 4×hackathon podiums, 2 national titles2022 – 2025
  • Intelligent AI Solutions (UK)
  • AIA Group
  • Ayala Group
  • Department of Health
  • Department of ICT
  • Philippine Science High School
  • TDCX AI Singapore
  • Philippine Genome Center
  • Carousell
§02 · datasheet

02 The 30-second version

Researcher turned forward-deployed AI engineer.

I started in genomics research at UP Diliman and the Philippine Genome Center, moved through data science at AIA and AI leadership at TDCX in Singapore, and now lead enterprise LLM delivery for a UK consultancy — working inside each client’s stack, data and constraints.

What I care about is the gap between “it works in a notebook” and“it runs reliably at scale” — async pipelines, evaluation, rate-limiting, governance and drift detection.

Full résumé (PDF)
PART NO. MC-AI-LEADREV 2026
Role
Forward-deployed AI engineer — currently Lead AI Engineer at Intelligent AI Solutions (UK), delivering into client environments
Shipped into
InsuranceBPO / customer serviceMedia monitoringTravelEducationPublic health & governmentGenomics research
Delivery
Discover → Scope → Ship → Harden → Hand over
Experience
6+ years in AI / data science · 30+ production-level projects
Focus
Enterprise RAGAgentic workflowsDocument intelligenceLLMOps
Core stack
PythonTypeScript · ReactFastAPIAzure OpenAIClaudeHybrid RAGMulti-agentMLflow · DVCDocker
Education
MS Data Science, Asian Institute of Management (2022) · BS Applied Physics (Instrumentation), University of the Philippines Diliman (2020)
Research
3 peer-reviewed publications — SARS-CoV-2 genomics, protein-interaction networks
Based
Philippines · Remote-first · UTC+8
Links
LinkedIn GitHub mcastro.dsci@gmail.com
STATUS: SHIPPING TO PRODUCTION
§03 · outcomes

03 Impact, by the numbers

Work measured in outcomes, not in demos.

Every figure comes from a real engagement — the company and project are on each tile, and projections are labelled as such.

95%

less FTE needed for Next-Best-Action model development

AIA Digital+ · up to 95% · end-to-end NBA data pipeline

● after  ·  ● effort removed

10–30×

document throughput — hours of turnaround cut to minutes

Document Intelligence Pipeline · batch / page / region concurrency

~40%cost saved

retrieval cost reduction via an LLM-driven RAG-skip gate

Hybrid RAG · Cohere rerank

20%+

proficient

faster time-to-proficiency with an LLM training simulator

TDCX AI · global BPO

2nd/ 123

teams at the UN Big Data Hackathon — Best Youth Team, Asia

International · 2023

25%

projected cut in claim-processing time from automated fraud screening

AIA Digital+ · fraud detection · projected

§04 · playbook

04 How I deploy

Embedded with the client, accountable for production.

The work lives in the gap between “it works in a notebook” and “it runs reliably at scale.” This is the loop I run on every engagement — each stage links to a case file where it happened.

  1. 01

    Discover

    Sit with the team that owns the process. Map the workflow, the data, and the constraints nobody wrote down — residency, access, spend.

    Leaves behindConstraint map

  2. 02

    Scope

    Cut the ask down to the narrowest MVP that moves a number the business already tracks.

    Leaves behindOne-page scope + success metric

  3. 03

    Ship

    Production on the client’s stack — their cloud, their identity provider, their network — not a demo on mine.

    Leaves behindDeployed service + IaC

  4. 04

    Harden

    Rate limits, retries, evaluations, drift checks and audit trails — before anyone calls it done.

    Leaves behindHardening log + eval suite

  5. 05

    Hand over

    Runbooks, training and dashboards so teams without ML engineers can operate and trust it.

    Leaves behindRunbook, dashboard, training

§05 · case files

05 Selected work

Nine systems. Architecture, code and outcomes for each.

Eight in production, one clearly-marked prototype. Every case file is a full write-up: the business problem, the system architecture, the engineering decisions with annotated code, and the outcomes.

01ProductionAgents

Document Intelligence Pipeline

Vision-LLM · 4-Phase Async Architecture

10–30×document throughput — hours of turnaround down to minutes

Converts complex multi-page insurance-style PDFs to semantic HTML using a 4-phase pipeline: Architect (layout analysis) → Specialist (adaptive extraction) → Builder (HTML assembly) → Verifier (quality checks). Achieves 10–30× speedup through three levels of async parallelism.

My role
Lead AI engineer · architecture + async pipeline
Stack
Gemini 3.0 FlashGPT-4.1Python asyncioPydanticDoclingChromaDB
Read the case file
FIG. 01 — SYSTEM ARCHITECTUREPDF batchmulti-pageArchitectlayout analysisSpecialistregion · asyncSpecialistregion · asyncSpecialistregion · asyncBuildersemantic HTMLVerifierquality gate
Map-reduce over agents: one Architect plans, a Specialist pool extracts in parallel, a Builder assembles, a Verifier gates quality.
03ProductionRAG

Enterprise RAG Chat Platform

Azure OpenAI · Multi-Strategy Retrieval · RBAC

3interchangeable retrieval strategies behind one governed, RBAC-gated platform

Production-grade RAG chat deployed on Azure Container Apps. Supports multiple retrieval strategies (semantic, vector, hybrid) with GPT-4V vision support for image-heavy documents, role-based access control, and per-user monthly credit limit enforcement.

My role
Lead engineer · platform architecture + governance
Stack
Azure OpenAI GPT-4oGPT-4V VisionPython / QuartAzure AI SearchAzure Container AppsCosmosDBAzure AD / RBACReact + Fluent UIBicep IaC
Read the case file
FIG. 03 — SYSTEM ARCHITECTUREUserAzure ADRBAC gatecredit limitsRewritequery plannerSemanticAI SearchVectorAI SearchHybridAI SearchGPT-4o+ GPT-4V vision
Every request passes an identity + credit gate, gets rewritten, then fans out to a pluggable retrieval approach before grounded generation.
04ProductionLLMOpsAgents

LLMOps Evaluation Platform

Agentic Evaluators · Ensemble Scoring · Experiment Tracking

0redeploys to change an evaluator — configs live in the database, code lives in git

End-to-end LLMOps platform for managing, running, and comparing AI evaluation workflows. Combines a Laravel API backend for config/auth with a Python FastAPI execution engine. Evaluators are pluggable Python modules; configs live in the database for zero-redeploy iteration.

My role
Lead engineer · evaluation architecture + scoring
Stack
OpenAI / ClaudePython FastAPILaravelMLflowDVCDocker
Read the case file
FIG. 04 — SYSTEM ARCHITECTUREConfig DBLaravel APIEval engineFastAPIJudgepersona AJudgepersona BJudgepersona CEnsembleTrust ScoreDriftmonitorMLflowDVC datasets
Agentic judges critique before they score; an ensemble weights them by confidence, and drift checks watch production.
06ProductionRAG

Hybrid RAG with Conversation Intelligence

Two-Stage Retrieval · Cohere Reranking · RAG Skip Logic

~40%retrieval cost cut — without giving up answer quality

Advanced RAG chatbot implementing two-stage retrieval (ChromaDB vector search → Cohere reranker), intelligent RAG skip (avoids redundant retrievals on follow-up questions), query reformulation from conversation context, and token-bounded conversation history management.

My role
AI engineer · retrieval strategy + session intelligence
Stack
OpenAI GPT-4oCohere RerankFastAPIChromaDBJWT Auth
Read the case file
FIG. 06 — SYSTEM ARCHITECTUREskipretrieveFollow-upmulti-turnRAG skip?LLM gateHistoryalready answeredReformulateresolve pronounsVector · 25ChromaDBRerank · 5CohereGPT-4ogrounded answer
A skip gate asks whether history already answers the follow-up; only if not does the two-stage retrieve → rerank path run.
InteractiveBurst-traffic labFire a burst of LLM calls through a naive rate limiter and the one I shipped — watch the 429s disappear. Runs in your browser.Open the lab →
02Production

Privacy-First Local RAG

Offline LLM Inference · Full MLOps Stack

0external API calls — fully offline RAG for regulated teams

  1. Ollama / Gemma 2 Inference
  2. ChromaDB Vector Store
  3. FastAPI Backend
  4. React + TypeScript Frontend
RAGLLMOpsRead the case file
05Production

Customer Service Agent Copilot

7-Endpoint Conversation API · GPT-4o-mini

7specialised copilot endpoints guiding live customer-service agents

  1. Journey Stage Classifier
  2. Resolution Path Mapper
  3. Scorecard Generator
  4. Outcome Forecaster
AgentsRead the case file
07Production

Automated Media Analysis Pipeline

Parameterised Notebook Execution · Parallel Orchestration

4analysis tracks run daily, fully automated and reproducible

  1. Papermill Orchestrator
  2. Parameterised Notebook Templates
  3. Sequential Execution Mode
  4. Parallel Execution Mode
AnalyticsRead the case file
08Production

Labor Market Skill Gap Dashboard

NLP · Curriculum Analytics · Interactive Streamlit

6dashboard views turning job-posting NLP into curriculum decisions

  1. NLP Skill Extraction Pipeline
  2. Taxonomy Alignment
  3. TF-IDF Gap Scorer
  4. NetworkX Co-occurrence Graph
AnalyticsRead the case file
09Prototype

AI-Powered Triage System

Claude · XML-Tagged Structured Output · Audit Trail

100%of triage decisions logged with reference IDs for audit

  1. Claude API Integration
  2. XML Output Parser
  3. MD5 Reference ID Generator
  4. CSV Audit Log
AgentsRead the case file

Metrics come from project-specific observation windows — happy to walk through any of them in an interview.

§ · studio

Studio · Fantasy Jam

After hours, I make puzzle games.

Fantasy Jam is my indie game studio. Its first game, InventoryDoku, is out now for Android on the Google Play Store.

InventoryDoku

Every item has a place.

Out now · Google Play

A Sudoku-style logic puzzle about shelving a night-shift warehouse delivery. Every rack has exactly one solution, and you can always reach it by logic alone. No guessing.

  • 60 campaign levels: each teaching at most one new rule, followed by practice racks before the next
  • Night Shift: endless generated racks in L, T and U shapes, a daily delivery and shareable stage codes
  • Hints that teach: a question first, then the reasoning step by step, with the pattern named
  • Free to play: out now for Android phones and tablets on Google Play
Get it on Google Play fantasyjam.io

Android · free to play · no access code needed

Official trailer · 2:00
  • A 6×6 rack mid-solve, with rule chips such as “Per shelf & bay” and “Can’t touch” above the board.
  • The home screen: an isometric warehouse with Campaign, Night Shift, Training and Employee of the Year.
  • A hint panel titled “Last slot left” walking through three numbered reasoning steps.
  • A Night Shift stage on a U-shaped rack with a gap in the middle.
  • A Hell-difficulty 9×9 rack with five rules.
  • A light pink rack skin with animal item icons.

Under the hood

Generated, then gated

Every level has exactly one solution

Levels are written as specs and generated, never hand-edited. The generator removes clues while a deduction stepper can still solve the board. A release gate then re-proves uniqueness with an independent search, grades difficulty and checks that every rule matters.

Deterministic on device

Same seed, same rack, on every phone

Night Shift builds stages on the phone, in time slices on the JS thread. Golden hashes prove Hermes and V8 produce identical racks. A 12-symbol share code rebuilds any stage and catches typos and swapped neighbours.

Cheat-resistant leaderboard

The server times the run, not the app

Vercel Functions, Upstash Redis and Clerk. Ranked runs start from a signed, single-use token. In Employee of the Year, the server keeps the seed and hands out racks one at a time, so none can be built ahead.

Launch, in code

Beta, trailer and ads from the real game

A closed-beta sign-up with an access-code admin and branded emails. Video ads recorded from live gameplay by a scripted finger that follows the hint engine. A two-minute trailer drawn frame by frame, with music and sound effects synthesized in code.

Expo SDK 57React NativeTypeScriptzustandVercel FunctionsUpstash RedisClerkAdMobRevenueCatVitest

§ · civic tech

Civic tech · Better LGU

Volunteer civic tech: local government, sourced and searchable.

I build and maintain independent, open-source community portals in the Better LGU network. Each one puts a town’s services, officials and public records in one place, with the source of every figure in view.

LiveLimay, BataanOpen source · MIT

BetterLimay

Community portal for the Municipality of Limay · builder & maintainer

A clearer way to find Limay government information, in English and Filipino. It isn’t the official LGU site and says so on every page: an independent, citizen-built guide that anyone can check against the original record.

  • Services: steps, requirements, fees and processing times from the Citizen’s Charter
  • Government: the 2025 elected officials and all 12 barangays with their officials and population
  • Transparency: COA-audited 2024 financial statements, plus procurement and infrastructure records
  • Statistics: census figures, DTI competitiveness rankings 2014–2024 and a UP NOAH flood hazard map
Visit betterlimay.org Source on GitHub
BetterLimay home page: “A clearer way to find Limay government information”, with a portal search box and popular services.
BetterLimay on a phone, with the emergency-hotline bar and portal search.

Provenance

Every fact names its source

Each record carries an exact public URL or a document cited precisely enough to request. A Python pipeline builds the data from a source register, and a validator checks it against JSON Schema in CI. Nothing is published without a source.

Privacy, in code

Rules the pipeline enforces and tests

Barangay officials appear with a middle initial only. Personal mobile numbers and SK members’ phone numbers are never published, only official office lines.

Quality gates

Every pull request has to pass

Lint, types, unit tests, Playwright end-to-end tests and axe accessibility checks. Lighthouse CI hard-gates accessibility, SEO and the script budget. Strict CSP and HSTS headers, with eval allowed only in the editor.

Built to hand over

Forkable, and editable by non-developers

Another town can rebrand the portal through config and data alone. Volunteers publish news through Decap CMS: saving a post opens a pull request, and merging deploys it. Cost to the people of Limay: zero pesos.

React 19ViteTypeScriptFuse.jsLeaflet · OSMRechartsPython pipelineVercelDecap CMSPlaywright + axeLighthouse CI

  • In progress

    BetterBalanga: the same approach for the City of Balanga, Bataan, listed in the directory as work in progress.

  • Directory

    Better LGU directory: contributor to the national list of community portals; maintainer of the Limay and Balanga entries.

  • Also

    Public-sector AI: Hack4Health champion with a health-fraud platform, DICT AI trainings, and a field note on augmenting caseworkers, not automating verdicts.

§06 · route map

06 Experience

From genomics research to enterprise LLMs.

Parallel lanes show where industry delivery, consulting and teaching ran at the same time.

2020202120222023202420252026B.1.1.7 genome paperUN Big Data · 2nd of 123USHER Hackathon · 3rdHack4Health · 1stHack-A-Street · 1stAI Builder Cup · Top 50INDUSTRYCONSULTINGTEACHINGEDUCATIONIntelligent AI Solutions · Lead AI Eng.CIIT & Eskwelabs · InstructorPSHS · Sr. DS ConsultantTDCX AI · DS ManagerAIA Digital+ · DSAIM · MS Data SciencePGCNOW
  1. Sep 2024 → PresentCurrent

    Lead AI Engineer · Tech Consultant

    Intelligent AI Solutions — UK enterprise AI consultancy · independent contractor

    • Led end-to-end development of enterprise LLM solutions: a media monitoring system for large-scale bias detection, and a RAG knowledge management system for a global travel agency.
    • Spearheaded AI research and production deployment: advanced RAG architectures, fine-tuning, scalable LLM implementations, agentic chatbots and automated workflow systems across multiple industries.
    Prototype → prodenterprise RAG + media-intelligence platforms shipped end-to-end
  2. Nov 2023 → PresentCurrent

    Data Science Instructor & Mentor

    CIIT & Eskwelabs

    • Instructor and mentor for workshops, bootcamps, hackathons and trainings on AI, No-Code AI and Data Science — for government organisations, corporate executives and professionals.
    Gov + C-suitecohorts trained on AI, no-code AI and data science
  3. Sep 2024 → Dec 2025

    Senior Data Science Consultant

    Philippine Science High School — Knowledge Innovation Division

    • Spearheaded data science initiatives — student performance management and data-driven curriculum enhancement for targeted academic interventions.
    System-widestudent-performance analytics for targeted interventions
  4. Aug 2023 → Sep 2024

    Global Data Science Manager

    TDCX AI — Singapore

    • Built an LLM-powered web app generating realistic simulation scenarios for agent training — cut time-to-proficiency by at least 20%.
    • Developed a forecasting model streamlining workforce management with predictive analytics and simulation in a user-friendly interface.
    20%+faster agent time-to-proficiency via an LLM training simulator
  5. Aug 2022 → Aug 2023

    Data Scientist

    AIA Digital+ — Asia-Pacific

    • Designed an end-to-end data pipeline for Next Best Action campaigns — reduced required FTE for NBA model development by up to 95%.
    • Built a fraud-detection solution to automate claim approvals — projected 25% reduction in processing time.
    95%less FTE for Next-Best-Action model development
  6. Sep 2020 → Mar 2021

    Senior Research Associate — Analytics & Visualization

    Philippine Genome Center

    • Reduced reporting times for genomic analysis of SARS-CoV-2 samples through custom data pipelines and analytical modules.
    • Co-authored peer-reviewed papers on emerging SARS-CoV-2 variants in the Philippines.
    3peer-reviewed papers from national SARS-CoV-2 genomic surveillance
§07 · system layers

07 Capability stack

Full-stack AI engineering, model to metal.

Mapped the way an AI system is actually built — and linked to the case files where each layer was put to work.

Systems I’ve wired AI into

Forward-deployed work is mostly integration: identity, data, runtime and the interfaces people actually use.

Identity & access

Data, search & storage

Runtime & infrastructure

Services & orchestration

Interfaces

  1. L5

    Models & GenAI

    Choosing, prompting and adapting the model for the job.

    • GPT-4o / o1
    • Anthropic Claude
    • Google Gemini
    • Azure OpenAI
    • Ollama · local LLMs
    • Hugging Face
    • LoRA fine-tuning
    • Prompt engineering
    • Stable Diffusion
    Proven inDocument Intelligence PipelinePrivacy-First Local RAGEnterprise RAG Chat PlatformLLMOps Evaluation PlatformCustomer Service Agent CopilotHybrid RAG with Conversation IntelligenceAI-Powered Triage System
  2. L4

    Retrieval & Knowledge

    Grounding answers in an organisation’s own documents.

    • RAG · Agentic RAG
    • Azure AI Search
    • Hybrid / semantic search
    • Cohere Rerank
    • ChromaDB · vector DBs
    • LlamaIndex
    • LangChain
    • OCR · Document AI
    Proven inDocument Intelligence PipelinePrivacy-First Local RAGEnterprise RAG Chat PlatformHybrid RAG with Conversation Intelligence
  3. L3

    Agents & Orchestration

    Multi-step pipelines, tools and the APIs that serve them.

    • Multi-agent systems
    • Tool use · function calling
    • MCP servers
    • FastAPI · Quart · Flask
    • Async Python
    • Pydantic
    • n8n · Zapier
    • Laravel · Node.js
    Proven inDocument Intelligence PipelineEnterprise RAG Chat PlatformLLMOps Evaluation PlatformCustomer Service Agent CopilotAutomated Media Analysis PipelineAI-Powered Triage System
  4. L2

    Data & Machine Learning

    Classical ML, NLP and the data underneath it.

    • PyTorch
    • scikit-learn
    • XGBoost · LightGBM
    • Time-series forecasting
    • NLP
    • Computer vision
    • PostgreSQL · DuckDB
    • CosmosDB · Redis
    • Pandas · NetworkX
    Proven inLabor Market Skill Gap DashboardAutomated Media Analysis PipelineLLMOps Evaluation PlatformEnterprise RAG Chat Platform
  5. L1

    Platform & LLMOps

    Shipping it, governing it, and knowing when it drifts.

    • Azure · AI Foundry · Container Apps
    • GCP Vertex AI
    • AWS
    • Azure AD · RBAC
    • Bicep IaC
    • Docker
    • GitHub Actions
    • MLflow · DVC
    • Evidently AI · W&B
    Proven inPrivacy-First Local RAGEnterprise RAG Chat PlatformLLMOps Evaluation Platform

Daily toolingClaude CodeCursorGitHub CopilotStreamlitPower BITableauPlotlyJupyter · Papermill

§08 · honours

08 Recognition

Four podiums. Two national titles. Peer-reviewed.

  1. 1stChampion

    ICSC Hack-A-Street

    Gravity model for retail site selection

    2025 · National · 13 teams

  2. 1stChampion

    DOH-DAP Hack4Health

    Fraud-detection platform for unified health data

    2024 · National · 100+ teams

  3. 2ndIntl. finalist

    UN Big Data Hackathon — Youth

    Best Youth Team, Asia

    2022 · International · 123 teams

  4. 3rd3rd place

    USHER Hackathon

    ML pipeline for building-damage early warning

    2022 · National

  5. Top50Grand finalist

    Google Cloud AI Builder Cup

    Impromptu: live AI-hosted podcasts that listeners can call into

    2026 · JAPAC · 21 countries

Publications

  1. [1]

    Detection and genome sequencing of SARS-CoV-2 belonging to the B.1.1.7 lineage in the Philippines

    Microbiology Resource Announcements · doi: 10.1128/MRA.00219-21 ↗

  2. [2]

    Genome sequencing and analysis of an emergent SARS-CoV-2 variant from Central Visayas, Philippines

    medRxiv · doi: 10.1101/2021.03.03.21252812 ↗

  3. [3]

    A directed minimum connected dominating set for protein-protein interaction networks

    Book chapter · isbn: 9780367814656 ↗

Speaking & training

  1. Government

    AI Technology Government Trainings

    Department of ICT · ICT Literacy and Competency Development Bureau

    Aug & Nov 2025
  2. Government

    BANT.AI: Revolutionising Fraud Detection with Unified Health Data

    Department of Health · 20th National Health Research Forum for Action

    Oct 2024
  3. Corporate

    Generative AI & Low Code for Corporate Teams

    Ayala Group of Companies

    Jun 2025 – present
  4. Corporate

    n8n: Automation Tools for E-Commerce

    Carousell Singapore · IT, Marketing & HR

    Aug 2025
  5. University

    Breaking into Tech Consulting Within 5 Years

    Asian Institute of Management · Career Services

    Mar 2025
  6. University

    Breaking Bias in AI

    FEU Institute of Technology · FEU Tech ACM

    Jun 2023
  7. University

    The Art of Data Storytelling and Data Visualization

    University of Santo Tomas · HORAIZON

    Mar 2023
§ · field notes

Writing

Field notes from production AI.

All writing

09 Contact

Let’s ship something real.

Email — fastest reply

Also taking select consulting engagements — enterprise RAG, agentic automation and AI enablement for teams. Scope a project →