Alejandro Fiestas

Computer science student at NYU

I'm from Lima and study computer science in New York, with psychology courses on the side, mostly neuroscience ones. I've worked two summers as a programming intern at a manufacturer in Lima. On my own time I work on projects like Ojo de Halcón, a tool for looking at Peru's public spending data, and on small EEG experiments.

Interests
Computer vision · Machine learning · Neuroscience
Based in
New York, NY · from Lima, Peru
Email me Download CV (PDF)

Selected work at Convertidora del Pacífico

Programming & systems intern, summers 2025 and 2026, Lima. A paper and corrugated packaging group of about 210 people. Each system is shown as the process it runs, with screenshots shared with the company's permission.

Month-end production costing, automated

C# · WPF · SQL Server

Problem. Costing for the paper and corrugated lines took about a month of manual work at each month-end, so finance was working months behind.

Built. A costing module that combines activity cost balances with production process data already in the ERP to compute the cost per product. It flags raw materials with abnormal costs and shows each product's list price next to its cost, so thin margins stand out.

Result. Runs in seconds instead of taking a month of manual work, so costing is now finished by the 15th of the following month. It surfaced manual costing errors across several past months, which finance corrected, and identified products that were being sold at very thin margins. Additionally, a companion sales analysis program I built also made thin margins visible, and caught a negative margin on stock bought at a high price during the pandemic.

Month-end costing, as a process
  1. Production process dataWhich products ran through which activities
  2. Activity cost balancesMonth-end balances per activity
  3. Cost per productAllocated across the lines
  4. Compare with list priceMargin per product
  5. FlagAbnormal raw-material costs and thin margins

Part of a .NET 9 / WPF inventory application modernized from the legacy ERP, with management-summary and data-integrity modules.

One month's costing, filtered to unreliable raw-material costs, then cost next to list price (sample data)
Difference analysis for August 2025 filtered to unreliable raw-material costs: 28 flagged lines, each with the computed cost, the cost recorded in the ERP, the difference in percent and the list price
Cost next to list price for a past month, with unreliable raw-material costs flagged (sample data)
Review before loading: flagged lines held back by type, and nothing generated until someone types CONFIRMAR (sample data)

From sales opportunity to approved credit line

.NET 9 · WPF · SQL Server · RUC lookup · SMTP

Problem. Sales reps requested new or expanded credit lines by email. Requests got lost, and reps weren't always told when a line was approved. The Head of Sales also couldn't track her team's progress, since there were no analytics on opportunities or potential clients. Reps had no way to keep track of their own tasks for each client opportunity.

Built. Sales CRMs for the paper and corrugated teams, with opportunity stages and team analytics for the sales manager. A credit-line platform where reps upload supporting documents in one place; existing clients load from the ERP and new clients autofill from Peru's tax-ID (RUC) lookup. Status updates go out by email with the request data consolidated. Both apps share one prospect table.

Result. Used by 50+ employees, about a quarter of the company. Users say it replaced email threads that often got lost.

Sales to credit. Email status updates go out at each change.
  1. CRM opportunityPipeline stages per team
  2. Closed-wonStarts the credit request
  3. Credit requestDocuments + tax-ID autofill
  4. Credit team reviewTracked in the app
  5. ApprovedProspect promoted to an ERP client

A new client enters the ERP's real client table only after approval, which keeps unapproved prospects out of it.

A new prospect from its RUC: the tax-ID lookup fills in the name and address (sample data)
Moving an opportunity to the next stage asks for its next task; then the deal is won and the sales manager's pipeline analytics open (sample data)
Credit request: client from the ERP, the required document attached, sent to evaluation (sample data)
Credit review: the resolution writes the approved line, rating and payment terms to the client's ERP record (sample data)

Meetings that turn into tracked tasks, with AI kept in check

React · TypeScript · Firebase · MCP server on Cloud Run · Claude API · Gemini API · Deepgram · Zoho API

Try the demoInvented data, simulated AI, nothing is saved

Problem. Formal meeting minutes and follow-up tasks across departments lived in documents and memory. The company requires an official "Acta" format to record its meetings.

Built. Seguimiento CDP, a PWA for meeting minutes and task tracking. Its MCP server lets people use the app from Claude, limited to what they can already see: they can ask what is overdue or what a meeting agreed on, or have Claude draft tasks. Claude cannot create or change a task; it only leaves proposals. A proposal drawn from a meeting or a document must quote the source word for word, one asked for in the chat is marked as a direct request, and nothing reaches the board until a person accepts it in the app. An audit log records who created, sent, or deleted each set of minutes, and when.

The platform also includes smart notes, a calendar, meeting scheduling, and a direct integration with the Zoho API: a change in Seguimiento CDP automatically updates the user's Zoho Calendar and Zoho Tasks. Meetings are transcribed with Deepgram and cleaned up with the Gemini Flash-Lite API. From the corrected transcript, either the Claude MCP server (using the user's own subscription) or the Claude API extracts a summary, a draft acta, and tasks. The agents use prompt caching and are tuned to keep token usage low and responses fast enough for a good user experience.

Result. Used by the board and by finance and planning leads, tracking 250+ tasks across 13 departments as of September 2026. Users can now produce actas easily and request a digital signature from the rest of the team, and tasks can be linked directly to the meeting they came from.

Meeting to task. Changes to the minutes are recorded in an audit log.
  1. MeetingFormal minutes
  2. TranscriptSource of truth
  3. AI proposes a taskMust quote the transcript verbatim
  4. Person approves or rejectsNothing lands unreviewed
  5. Task boardAcross 13 departments
Proposed tasks quoting the transcript, accepted one by one (demo data)
The Logística board in the demo: tasks to do, in progress and done, with owners, due dates and priorities, grouped by project, and the agent's proposed grouping waiting for approval
An area's task board, grouped by project (demo data)
Seguimiento CDP on a phone, recording a meeting: a pause button, a running timer, the Concluir reunión button and a transcript section still empty; the organiser's name on the meeting card is pixelated
Recording a meeting from the phone
Asking Claude for a task through the MCP server, then accepting its proposal in the app (other people's proposals pixelated)
The agent screen of Seguimiento CDP on a phone: suggested questions about overdue work, the last meeting and an area's week, and a note that it changes nothing on its own; its task proposals stay suggestions until someone accepts them
The in-app agent, which changes nothing on its own

Modernizing the legacy systems

C# · .NET 9 · WPF · ASP.NET Core · SQL Server

Problem. The ERP ran on Centura (Gupta) programs from another era, with the deployment and access habits that came with them.

Built. A C# parser that converts legacy source into an intermediate format for AI-assisted translation to .NET 9 / WPF. Using it, I migrated several programs the business depends on: Stock Corrugado and Stock Papeles (inventory levels; one program in corrugated and paper versions), Movimientos de Inventario (the kardex, or stock movement ledger), Reporte de Ventas (sales reporting), and Seguimiento de Compras Locales (local purchase tracking). I also started a new ERP launcher to replace the Centura one, with per-user database logins, since handed off to a colleague. Alongside the migration, I built a set of new applications the company needed and didn't have. None of the systems listed below is a translation of a legacy program; each was built from scratch around the ERP.

Result. Four legacy programs migrated and the migration ongoing; an auto-updater now runs in three production desktop apps, with internal patterns packaged as Claude Code plugins for reuse.

Reporte de Ventas after the migration: classification, monthly trend, top articles (sample data)
Productos Diferidos: what was invoiced but not yet dispatched at a cutoff date, by month of issue, with totals in soles and dollars, the number of invoices with an open balance and their average age
Productos Diferidos, a new application: invoiced but not yet dispatched, by month of issue at a cutoff date (sample data)
Dispatch board, Entregas comprobadas: orders ready to schedule, routes built and routes on the road. Opening an order shows its lines; opening a route shows its stops and draws it on the map (sample data)
The same board's map view: it follows each truck across Lima in turn, with the fleet listed beside it and the ones without GPS marked (sample data)
Trucks on the road (sample data)
The dispatch board on a phone, one route: its truck on the map of Lima between the plant and its stop, the driver, the load and the share of capacity used, and the stop, Alicorp in Callao, delivered
A route and its stops (sample data)
The phone as a remote for the wall screen named DEMO, which is showing Entregas comprobadas: arrow keys, OK, zoom and Atrás
The phone as a remote for a wall screen

New applications, not migrations

  • Dispatch board. Real-time board for warehouse wall screens and phones: orders ready, scheduled, en route, plus a live GPS fleet map. A phone can act as a remote for a wall screen. One shared snapshot keeps load at 240 queries per hour regardless of screen count.
  • Deferred products. Tracks what was invoiced but not yet dispatched, which happens when a client changes an order or agrees to take slightly less, among other reasons. Shows the open balance at a cutoff date, by month of issue.
  • Database inconsistency detector. Read-only checks that surface data-integrity problems in the ERP database.
  • Payment tracker. Projects monthly obligations to international suppliers against a spending cap.
  • Import tracking app. Tracks incoming shipments for the purchasing team.

Projects

Plushcut

Design and production tool for custom sublimated plush, launching in Lima. In development since September 2026.

React 19 · TypeScript · Vite · Konva · three.js · WebGPU · ONNX Runtime · Clipper · Python (FastAPI) · IndexedDB

Custom sublimated plush is made by printing a photo on polyester, cutting along the silhouette, sewing front to back and stuffing it. Plushcut does everything between the photo and the print shop. It cuts the subject out of the photo, traces the silhouette, and builds the pieces at real size with seam allowance and bleed — front and back for a pillow, gores for a ball — then shows the sewn result in 3D, packs the pieces onto the fabric sheet the shop prints on, and exports what the shop needs: real-size PDF, PNG, JPG or TIFF, in RGB or in CMYK through the shop's own colour profile, plus a DXF or SVG cut file for a laser cutter and an A4 paper pattern for cutting by hand. It also works out what a sheet costs and what margin is left per product. The interface is in Spanish, it runs offline in the browser, and the geometry, imaging and 3D live in separate packages so a customer-facing ordering page can reuse them.

Decisions worth showing

  • No production number is hard-coded. Seam allowance, bleed, sheet size and the rest resolve in layers — fabric, then production profile, then project, then piece — and the editor shows which layer each number came from.
  • Nothing derived is stored. A design keeps only its sources; cut, stitch and bleed lines are recomputed every time. Sheet layouts are saved as rules rather than coordinates, so changing the sheet size or the seam allowance re-solves the layout instead of leaving stale positions.
  • Seam allowance is a real polygon offset, not a scaled-up silhouette. Scaling changes the shape; sewing needs the same distance from the edge all the way round.
  • Two geometry problems were worth the trouble: predicting what a flat outline will look like once it is stuffed, after a first attempt creased down the middle; and flattening the curved panels of a ball so the edges that get sewn together still match, within 6%.
  • Packing the sheet exactly was too slow — 70 to 640 ms for a single pair of pieces. A coarse grid now proposes positions and the real geometry only checks those, which fits irregular shapes on a sheet in under a second.
  • The cut file is simplified but the print is not. Offsets and traced photo outlines leave clusters of sub-millimetre segments that make some cutter controllers pause at every point; only the vectors sent to the cutter are simplified, within a tested tolerance.
  • It runs offline in the browser. Photos, projects and the shape library live in browser storage and move as files; background removal and upscaling run locally on the GPU from ONNX models, so a customer's photo never has to leave the machine.
A pet photo becomes two sewable pieces: crop, background removed, silhouette traced, cut and stitch lines at real size
Size, border and a name on the front piece; below, the same design against a coin, a phone, a bottle and an A4 sheet
Filling a 160 × 100 cm sheet: pieces nest automatically and the free space takes smaller ones, while fabric used and cost per sheet keep up (cost estimates)
The sewn result in 3D before anything is printed, then beside a phone and a bottle at the same scale

Ojo de Halcón

Public-procurement data for Peru. Pre-release; public release coming.

Python · DuckDB · SQL · Parquet · TypeScript · Next.js · DuckDB-WASM · Cloudflare Pages, R2, D1

Open the pre-releaseClosed beta, by invitation only

A platform that makes Peru's public procurement data explorable for citizens, journalists, and researchers: 16 SQL risk indicators, district-level views of public works, and an explorer for deeper analysis. Indicators are signals for review, not accusations.

139
months of official procurement data
868,773
procurement processes from 3,162 entities
18.3M
rows in the national taxpayer registry, cross-referenced
191,177
public works from the Comptroller's open INFOBRAS data
Data flow. No application server: queries run in the visitor's browser.
  1. Official open dataProcurement, taxpayer registry, public works
  2. Python ingestionBulk loading, about 7 minutes
  3. DuckDB warehouse35+ data-contract checks gate publishing
  4. Parquet on Cloudflare R2Static files
  5. Browser queriesDuckDB-WASM

Decisions worth showing

  • Caught a size bias: an early indicator fired on 0.45% of tenders with 2–3 bidders but 39.2% of those with 61 or more. It was flagging big tenders. Measured, corrected.
  • Refused a merge that looked like more coverage. Across 3,838 works dated in both sources, not one completion date matched; merging would have doubled flagged works (1,135 vs. 558). The second source also contradicts itself: 42.9% of works marked finished report under 100% progress.
  • Personal data is never read. The extractor takes 34 of 97 columns and never loads supervisors' or residents' names (Law 29733). A deny-by-default personal-data check is one of the 35+ contract checks.
  • Hand-checked project geolocation on 305 cases, reaching 99.3% accuracy in the final round.
  • Found a slowdown from row-by-row loading (6 to 14.5 minutes) and fixed it with bulk loading.
  • Every published figure shows its coverage. Figures that didn't reproduce were corrected and documented.
Ojo de Halcón district page for Villa El Salvador: population, this year's municipal budget and how much has been spent, and a note that an active indicator is a question about a public record, not a conclusion
Ojo de Halcón home page: the headline, a short description and the district search box
Registry figures for works and districts, and the altimeter that explains the levels of detail and what a signal means
The explorer: entities table with amounts, process counts and active indicators, plus filters and the altimeter

Qatu

Offline-first point of sale and inventory for Peruvian corner stores. In development; tested by me and friends, not yet used by stores.

React 19 · TypeScript · Vite · Tailwind · Firebase · PWA · GitHub Actions

A mobile-first alternative to an ERP for bodegas: a point of sale that connects to a barcode scanner so each scan is instant, with the phone camera as an alternative, and takes cash, Yape/Plin, or card; inventory with weighted-average cost, low-stock alerts, and a full movement history per product; daily cash closing, expenses, store credit (fiados), and supplier debts; and a shared barcode catalog with Open Food Facts as a fallback. The original idea behind the app was to let small businesses reach competitive wholesale prices through group buys.

Decisions worth showing

  • A sale never waits on the network. Stores lose signal mid-sale, so sales save locally and sync later.
  • Stock and cost can't be typed in directly. Stock moves only through sales, restocks, and adjustments; cost only through the weighted average.
  • Past sales are never rewritten, so old reports stay true even after a product is deleted.
  • Each day's cash closing is saved under its business date; a recount corrects the day instead of creating a contradictory record.
  • CI blocks merges unless the app builds, unit tests pass, Cloud Functions build, and Firestore security-rule tests pass.
New product: scan, catalog autofill, cost and price, kardex
Sale by scanning: three products, cash, registered offline
Sale by tapping the products: search, tap, cash, registered offline
Qatu inventory on a desktop screen: low-stock warning, total stock value, and the product list with prices and units
Inventory on desktop (demo data)

Sumay

Personal planner, installed as a PWA. Live since February 2026; used by me and a few testers.

React 19 · TypeScript · Vite · Tailwind · Firebase (Auth, Firestore, Cloud Functions) · PWA · Gemini API

A planner built for my own days. The dashboard shows the current event on a focus ring or the whole day on a 24-hour wheel, with the agenda below. The calendar stores recurrence as RFC 5545 rules; tasks are events with deadlines; habits are counted or timed, with streaks and stats; trips are day-by-day itineraries with an invite link and an AI planner. It imports Google Calendar and can push changes back, and reminders arrive as web push even when the app is closed. Five languages, 17 themes.

Decisions worth showing

  • Recurrence is stored as a standard RRULE rather than a custom schema, so daily, weekly, monthly and "every N" series, end dates, and edits to one occurrence all follow one set of rules, with 18 tests covering recurrence and daylight-saving changes.
  • Reminders are scheduled server-side: a Cloud Function places each one in a Cloud Tasks queue, so a push arrives on time whether or not the app is open.
  • Habit timers keep running across pages, so a timed habit started on the list keeps counting on the dashboard.
Habits: complete the morning list, start the plank timer, watch it on the dashboard
Sumay calendar, month view, with dots marking each day's events and tasks
Month view; dots are the day's events and tasks
A trip plan for Mancora in August 2026 with an invite button and no activities yet
A trip plan, shareable by invite

Cooliving

Co-living platform for Latin America, built for MBA students who pitched it as a business idea. May–July 2026.

React 18 · Vite · Firebase (Auth, Firestore, Storage, Cloud Functions, Hosting) · Leaflet · Gemini API · PWA

A co-living platform for Latin America, Spanish-first with an English version. People looking for a room answer an affinity test on schedule, cleanliness, noise, guests, work style, social life, pets and smoking; listings that pass hard filters are ranked by a 0–100 compatibility score. Hosts publish rooms with photos and a map pin. Roommates can form a group and apply for a whole apartment together, with invites and votes handled by Cloud Functions. Contracts are generated from Lima and Bogotá templates. Gemini runs a chat agent that fills in missing compatibility answers, explains the top matches, mediates household incidents, and drafts a household report for hosts.

Decisions worth showing

  • The affinity score runs in the browser and in a Cloud Function, and the server copy is kept byte-for-byte identical, so a group's score comes out the same wherever it's computed.
  • The Gemini key is stored as a Cloud Functions secret. The browser never calls the model; every AI feature goes through a server function.
  • Identity verification sits behind a provider interface: the app ships with a mock, an adapter for one vendor (Truora) is written, and the webhook must check the vendor's signature before writing anything. It isn't enforced yet.
A host builds a listing: photos, map pin, amenities and rooms (demo)
A seeker takes the affinity test, then sees a ranked room (demo)
Cooliving room search: two listings in Bogotá, a full apartment in Chapinero Alto and a loft in Teusaquillo, with photos, spots available and monthly prices
Room search in Bogotá (demo listings)
The Cooliving landing page: the headline Porque convivir es Cool con Cooliving, a city and date search, and a shortlist of roommates ranked by affinity
The landing page, with a roommate shortlist ranked by affinity

EEG experiments

Ongoing, incomplete. Where computer science meets the psychology minor.

BioAmp Pill · Python

Small projects that record brain signals with a BioAmp Pill and use them to control a computer. Collected my own recordings and trained a model to improve detection accuracy. Unfinished, and listed here because it's the work I keep coming back to.

Automatic license plate recognition

March–October 2021

Python · OpenCV

Detects Peruvian license plates in security-camera footage and logs vehicles entering and leaving in real time. Trained a Haar cascade on real footage from a residential camera. Plate-reading (OCR) accuracy was limited by the low-resolution footage.

The pipeline on screen: labelled samples, detection at the gate and the count, plate crops filling the output folder, character segmentation
Five small frames from the motion detection steps: the original camera frame, the difference between frames, grayscale with blur, the threshold mask, and the dilated mask
Motion detection: original, frame difference, grayscale and blur, threshold, dilation
A car seen from behind with a blue motion contour around it and a green box where the Haar cascade found the plate; the plate is pixelated
The Haar cascade finds the plate inside the moving contour (plate blurred)
Five crops of the same plate through the preprocessing steps: original, black-hat, Canny edges, Hough lines, and four coloured dots at the detected corners; the characters are pixelated
Plate preprocessing: original, black-hat, Canny edges, Hough lines, corner detection

How I work

I build with AI coding agents every day and write internal tooling for them, but I make the design decisions, measure the results, and verify the outputs myself. I run tests and review the code before deploying or committing to a design solution.

Most of what I have built is for Peruvian contexts: a manufacturer's operations, corner-store owners, and public spending data. My first computer vision project was in 2021, and it is what sparked my interest in computer science, and in computer vision specifically. Outside CS I am drawn to psychology and neuroscience.

I work well independently but adapt easily to different kinds of teams. I like discussing problems, their limitations, and possible solutions. During my internship, what I enjoyed most was meeting the people who actually used the software I was building; it connected me with people across many areas of the company. I also like cats.

Education

New York University, College of Arts & Science · expected May 2028

B.A. in Computer Science, intended minor in Psychology. Dean's List 2024–25.

Basic Algorithms · Operating Systems · Computer Systems Organization · Data Management and Analysis · Software Engineering · Introduction to Computer Security (in progress)

Newton College, Lima · IB Diploma, 2024

Extended essay: how CNN and KNN accuracy changes with training-set size when detecting brain tumors in MRI scans.

Skills

Languages
Python, TypeScript, JavaScript, C#, SQL
Frameworks & data
React, Next.js, .NET 9 (ASP.NET Core, WPF), DuckDB, Parquet, OpenCV
Cloud & tools
Firebase, Google Cloud Run, Cloudflare (Pages, R2, D1), SQL Server, Git, GitHub Actions
AI-assisted development
Claude Code, Model Context Protocol (MCP)
Spoken
Spanish (native), English (fluent)

Available for a 12‑week internship, May–August 2027.

afiestaspublic@gmail.com Download CV (PDF)