Document AIInference systemsBackend platformsPortfolio 2026

Applied AI and the systems around it

I'm Manuel, a lead software engineer with more than six years of experience building backend, data, and financial systems. I've led document-extraction platforms, designed multi-tenant reporting infrastructure, and debugged 500-billion-parameter model serving on Blackwell GPUs. I work across architecture, implementation, evaluation, and production—and I still spend most of my time writing code.

What I work on
Document extraction with source evidence
Multi-tenant APIs and data platforms
Self-hosted inference and model serving
Financial and cryptographic systems
6+ years
shipping backend, financial, and applied-AI systems
34% → 94%
evaluation accuracy on bilingual audited statements
504B MoE
served across 8 RTX PRO 6000 Blackwell GPUs
Preview CV ->Download CV ↓Engineering notes ->GitHub ->LinkedIn ->

Selected work

Production systems and technical investigations I led or built. Each example states what I owned, the constraint that mattered, and the result.

01
Document AI / financial extraction

Audited financial statement extraction

I led the technical work for a Saudi compliance platform and owned its OCR and extraction service. It converts Arabic and English audited statements into structured financial data while retaining source evidence for every value.

34% → 94% eval accuracy

The service compares two independent extraction passes, then checks their output against accounting identities, section totals, and cross-year consistency. Values that fail validation go to human review instead of entering the system silently. Evaluation accuracy improved from 34% to 94%; a 22-page filing processes in 97 seconds for $0.30–$0.42.

TypeScriptPythonFastAPIPostgreSQLLLM visionOCR
02
Inference / vLLM / Blackwell

GLM-5.2-504B on RTX PRO 6000

I served a 504-billion-parameter MoE checkpoint on eight RTX PRO 6000 Blackwell GPUs and traced a deterministic repetition failure in vLLM's sparse-attention path.

240K cold context

The weights and sparse kernel were valid. The checkpoint exposed layer behavior through indexer_types, while vLLM expected index_topk_pattern; because that field was null, the indexer ran on the wrong layers. Deriving the pattern restored correct output. Decode-context parallelism and 8,192-token chunked prefill then enabled cold prompts up to 240K tokens.

vLLMCUDANCCLB12XNVFP4GCP
Read the investigation ->
03
Multi-tenant document reporting

Traceable construction reporting

I lead the architecture and OCR work for a construction-reporting platform that ingests daily-report PDFs and SharePoint files across owner, contractor, and subcontractor organizations.

7 traceable workflows

I designed the raw-to-normalized data model, source provenance, validation flags, and reporting APIs for manpower, equipment, costs, and daily briefs. Every normalized value keeps its source file, page, section, raw label, and validation state. Project and tenant boundaries are enforced through scoped credentials and database policies.

FastAPIPostgreSQLPydanticReactPDF workers
04
Disaster response / search

Venezuel.help

I built Venezuel.help in 24 hours after the June 24 La Guaira earthquake. It combines scattered hospital, missing-person, and aid records so families can search them from one place.

Built in 24 hours

The ingestion pipeline reads web pages, PDFs, and photographed lists. OCR and structured extraction turn them into person records, while cédula matching and name normalization reduce duplicates. Scheduled checks revisit sources as statuses change. The public site also maps building damage and aid centers.

Mistral OCRGemmaCerebrasMapsScheduled ingestion
Open the live site ->
05
Reverse engineering / model runtime

Gemini Nano from Chrome

I used Ghidra, Frida, and runtime tracing to recover Gemini Nano's model architecture and INT4 weight layout from Chrome's 3.98GB binary.

3.98GB binary / ongoing

The investigation traces graph construction, weight binding, and the FP32-to-FP16 side-table writers. One key finding is that model reload only rebinds cached tensors; packing runs during initial graph construction. That changes where hooks must be attached and explains why the obvious reload traces never reached the writers.

GhidraFridaWindowsINT4FP16Runtime tracing

Other work

Open-source projects, client systems, and focused experiments. Public entries link to code or a live product.

  • Threshold consent verification — built a Rust/WASM backend using 7-of-10 threshold decryption, BLS aggregate signatures, and Merkle commitments; shipped SDKs in Dart, TypeScript, Java, and Rust.
  • Aave position protection — monitored 84 smart-wallet positions for 18 months, recalculated risk every five minutes, and completed the production run with no managed position liquidated.
  • WDK browser wallet — public Manifest V3 starter with a background-only key boundary, AES-256-GCM vault, origin-scoped dApp permissions, transaction decoding, and 290 tests.
  • DownloadToMarkdown — document-conversion service for scanned PDFs, images, and DOCX files, with large uploads routed through a dedicated TLS origin. ->
  • Aave V3 data API — self-updating reserve data for 13 networks and 217 assets, queried with JSON-RPC and Multicall3 and published through GitHub Actions. ->
  • Colab inference — turns a free Colab T4 session into a local OpenAI-compatible endpoint; Chrome DevTools Protocol automation handles deployment. ->
  • AI work ledger — Zig, native Win32, and WSL tooling for recording AI-assisted work without screenshots or keystroke logging.
  • HandyGems — rebuilt a client's Electron voice assistant in .NET 8 with realtime Gemini Live audio, vision input, and working agent tools.
  • MedAgenda — offline-first patient records for Venezuelan clinics, with OCR running on-device so documents do not need to leave the phone.
  • DermaCare — scheduling and patient records used daily by a dermatologist, including ICD-11 coding.
  • OSB fantasy sports — NestJS backend with Stripe payments and realtime WebSocket scoring; the MVP led to a year-long client engagement.
  • Trading strategy harness — leveraged-strategy backtesting with an end-to-end audit for bias in the evaluation path.
  • Quantum-seeded wallet — conditioned raw IBM quantum-hardware bitstreams with NIST SP 800-22 tests and Toeplitz extraction before key generation.
  • Ethereum RPC gateway — Go service with adaptive endpoint selection, proxy rotation, and Prometheus metrics for unreliable providers.
  • Zero LSP — diagnostics, hover, completion, and semantic tokens for a programming language I'm building.
  • PR review router — self-hosted service that sends each GitHub pull request to existing AI review subscriptions and collects the results in an admin dashboard.

About

My foundation is backend and financial systems. Over the last several years, my work has moved closer to the model layer: OCR, structured extraction, agent workflows, self-hosted inference, and runtime debugging.

I work best on small teams where I can talk to users, define the system, and implement the critical path. I work remotely from UTC−4 in English and Spanish.