Projects and research

Work

Selected software I have designed and built, and research I have contributed to. Each project shows its current stage; the studio engagement is confidential and is described at the level of capability.

Code: github.com/malvarado-techProfile: linkedin.com/in/malvaradogzz

  1. ProductionSince July 2026

    Finance and administration platform for a creative studio

    Operations Consultant (Finance, Administration & Systems)

    The studio had no financial baselines or per-project figures, and its payments, payroll records and collections lived in separate spreadsheets. The platform brings them into one ledger, so the directors can see where the studio stands and payments go out on a weekly schedule, with approvals.

    Designed and shipped the studio's internal finance and administration platform, in production on AWS since July 2026: ledger, collections and retainers, payment requests with approvals and a weekly payment run, quotations with a costing engine, payroll records, invoice tracking with fiscal checks and director dashboards.

    Changes reach production only after automated integration tests against a real database, so the platform can be extended every week without interrupting the team's work.

    The platform also hosts a daily ingestion pipeline from the Meta Marketing API into Postgres and Google Sheets, and an LLM assistant that calls the platform's own tools under the user's role permissions.

    Confidential client. Screenshots use fictional demo data.

    Stack

    • React
    • Tailwind
    • Node
    • Postgres
    • AWS
    • GitHub Actions
    • Vitest
    • Meta Marketing API
    • LLM APIs

    Director view: cash position, runway, receivables, monthly margin and recurring revenue on one screen

  2. Client engagementOct 2025 to Aug 2026

    biorisk.ai: AI service layer for a biosafety risk-assessment product

    AI Engineer & Technical Project Manager

    Biosafety assessments have to be defensible, so the product could not rely on a model answering from memory. Every answer is built from a curated set of authoritative sources and cites them, and the service runs self-contained and supervised, so assessors can check each recommendation.

    Designed and built the knowledge base, data pipeline and AI service layer in Python on AWS: a retrieval-augmented generation (RAG) service with semantic search and citation tracking on every answer, on a serverless architecture (Lambda, Bedrock) with no dependence on third-party services.

    Designed model routing by assessment type, with automatic failover on a refusal.

    Integrated the insights pipeline into the partner's Django backend behind a feature flag, with rate limits and a rollback plan, validated against real production assessments.

    Client engagement, delivered August 2026.

    Stack

    • Python
    • AWS Lambda
    • Amazon Bedrock
    • RAG
    • Semantic search
    • Django
    Architecture of the AI service layer (diagram)Authoritative sources go through curation and quality review, then extraction and chunking, into embeddings and a knowledge base with semantic search. Retrieval feeds an insights agent and a conversational agent, which share model routing with failover and return answers with citations. Everything after the sources runs serverless and self-contained.Authoritative sourcesServerless,self-containedCuration and quality reviewExtraction and chunkingEmbeddings and knowledge basesemantic searchRetrievalInsightsagentConversationalagentModel routing with failoverAnswers with citations
    Architecture of the AI service layer (diagram)
  3. Proof of conceptMar to Apr 2026

    IBBIS Screen: web interface for an open-source DNA-synthesis screening tool

    Self-initiated build

    IBBIS's Common Mechanism runs from the command line, which keeps it out of reach of many of the technicians at synthesis providers who are meant to use it. A web interface removes that barrier and leaves the screening itself to IBBIS's tool.

    Built and deployed a web interface for IBBIS's open-source Common Mechanism DNA-synthesis screening tool, so that synthesis providers and technicians can run a screen without the command line or bioinformatics expertise.

    Proof of concept, released as open source under the MIT license. I presented it to the IBBIS team through an email exchange and a meeting, where they asked about the design rationale and for my input on bringing the tool to wider audiences, their planned next step. The screenshots and the linked demo come from its demo mode, which runs in the browser with sample data.

    Stack

    • Next.js
    • FastAPI
    • Docker
    • Google Cloud

    Landing page: what the interface does and how a screen runs

  4. PilotSince May 2026

    Real-time AI translation platform for events

    Self-initiated build

    Regional conferences such as EAGx could not offer live interpretation because of its cost, so part of the audience followed talks in a second language. The platform lets an event offer its talks in several languages on attendees' own phones, at a fraction of the cost.

    Designed and built a real-time AI translation platform on Next.js and TypeScript: the speaker's audio goes to a realtime speech model over WebRTC and the translated audio is republished through a self-hosted LiveKit server to listeners' phones, with an operator dashboard and a listener web page.

    Credentials stay on the server: each session receives short-lived tokens minted server-side and scoped to its role, so the browser never holds an API key and a listener can receive the translation but never publish to the room.

    A second provider's realtime speech model runs as a parallel A/B engine, and both engines are compared on the same audio across logged sessions, with a written decision record.

    The linked demo replays a 1962 public-domain recording through the listener and operator views, with translated captions and AI-generated voices, in the browser and without keys.

    Stack

    • Next.js
    • TypeScript
    • WebRTC
    • LiveKit (self-hosted)

    Replay of a 1962 recording with translated captions and AI-generated voices in Spanish and Portuguese

  5. Proof of conceptMay 2026 snapshot

    Field audit platform for an environmental, health and safety consultancy

    Design and development

    Inspectors wrote their field notes and then retyped them into each report, and expiry dates were tracked by hand. The platform captures a site visit once and produces the deliverables from it, and it flags commitments before they expire.

    A proof of concept for workplace safety and compliance audits, shown as a May 2026 snapshot. Inspectors record a site visit once (findings by area with priority and regulatory reference, extinguisher inventory, environmental checks, ergonomic questionnaires) and the platform produces the deliverables from that single capture.

    A team dashboard surfaces recent audits, overdue extinguishers and open high-priority findings; an expiry tracker sorts commitments into overdue, critical and upcoming.

    Stack

    • React
    • Vite
    • Express
    • SQLite

    Team dashboard: recent audits, critical expiries and open high-priority findings (demo data)

  6. Functional demoMay to June 2026

    Service-report platform for an industrial water-treatment company

    Design and development

    Field advisors service cooling towers, boilers and softeners at industrial plants every month. Readings were written by hand on site, re-keyed into a spreadsheet, photos were pasted into cells and the sheet was printed for signatures; by the company's own estimate each report took a full day, and nothing flagged a reading outside its ideal range.

    The advisor captures make-up water, chemical readings and the physical inspection for every unit on one screen, and each reading is checked against that unit's spec range as it is entered and flagged when it falls outside. Sign-off happens on screen, one click produces a bilingual PDF and a shareable link for the customer, and every unit keeps a trend chart across visits. Equipment types and vendor ranges live in editable registries, so the catalog changes without a code change.

    Functional demo, deployed for the client's evaluation. The linked static demo runs in the browser with fictional customers, equipment and readings.

    Stack

    • React
    • Vite
    • Tailwind
    • Express
    • SQLite
    • Puppeteer

    Operations dashboard: visits, readings in range, units out of spec and open visits (demo data)

  7. Advanced demo2024 to 2026

    Origami Hub: equipment rental and inventory platform for a creative studio

    Operations systems designer

    The studio wanted to organise how it lent equipment to its own productions and to put its spare equipment to work by renting it to outside productions, as an additional source of income for the space. The kiosk was designed to bring a check-out from three to five minutes to under fifteen seconds, and the public site to take rentals.

    Started as an internal kiosk inventory tool for a creative studio and grew into a dual-platform equipment-rental solution: an internal kiosk app with hardware barcode-scanner check-out and check-in, and a public rental site with real-time availability and automatic conflict detection.

    Taken to an advanced functional demo deployed for evaluation; not in daily use. The linked static demo shows the public landing page and catalog with fictional equipment and prices.

    Stack

    • React
    • Vite
    • Express
    • Postgres
    • Next.js

    Public rental site: home page

  8. Research2025

    Research on LLM evaluation

    First author and co-author

    Most LLM leaderboards rest on a single run, and translated benchmarks can hide language-specific reasoning errors. The two preprints test both assumptions: how many repetitions a reliable ranking needs, and how models reason on problems written natively in Spanish.

    • Do Repetitions Matter? Strengthening Reliability in LLM Evaluations

      First author · arXiv preprint, September 2025

      Re-evaluating eight models on AI4Math with three runs each: under a single run, 10 of 12 evaluation slices inverted at least one pairwise ranking, and a second run removed about 83% of those inversions.

      arXiv:2509.24086
    • AI4Math: A Native Spanish Benchmark for University-Level Mathematical Reasoning in Large Language Models

      Co-author · arXiv preprint, May 2025

      105 original university-level problems authored natively in Spanish across seven domains, each with a step-by-step human solution; six models evaluated in Spanish and English, zero-shot and with chain of thought.

      arXiv:2505.18978