Open to full-time roles · starting January 2027

David Meltonyan.
I build ML that ships
and the backends behind it.

Software engineer focused on machine learning and backend systems. Three internships putting models into real workflows, from insurance underwriting to retail analytics. CS at Sacramento State, graduating December 2026.

Based in
Sacramento, CA
Degree
B.S. CS · Dec 2026
Next
M.S. CS · Georgia Tech
Focus
ML · Backend · Data

Featured work

Click through to the code, or try the live demo.

Machine learning · Full stack · Live

Fraud Triage

Fraud models are easy to train and hard to act on. This system turns a LightGBM fraud probability into a decision (approve, send to an analyst, or block) by comparing expected dollar cost under a fixed daily review budget, then gives analysts a review app with the reasons behind each score.

  • Leak-free by test. Time-based splits only, and a test proves no feature can see the future.
  • Calibrated scores. Platt scaling on a held-out month so probability × amount means real dollars.
  • Runs on free hosting. Scoring is offline; the API serves precomputed results and a test checks it never imports the model.
PythonLightGBMSHAPDuckDBFastAPIPostgreSQLNext.jsPlaywright
Total cost on 3 held-out months, 30 reviews/day
Cost-based policy$12,762
Review highest probability$37,964
66%lower total cost than the standard approach
0.838test PR-AUC vs 0.376 logistic baseline
93.3%of fraud caught reviewing the riskiest 1%
1.85Mtransactions, time-split train / calibrate / test
Backend · LLM infrastructure

CostMelt

A drop-in HTTP proxy that sits between an app and LLM providers and cuts inference cost without client code changes: it classifies prompt complexity and routes to the cheapest capable model, serves semantically similar prompts from cache, compresses prompts, and micro-batches requests.

  • Routing engine with a complexity classifier across OpenAI, Anthropic, Groq and DeepSeek.
  • Semantic cache on pgvector embeddings with Redis TTL and LRU eviction.
  • Per-request cost attribution surfaced in a Next.js analytics dashboard.
PythonFastAPIRedisPostgres + pgvectorNext.jsDocker Compose
Request path
  1. 1 · Cache similar prompt seen? return instantly
  2. 2 · Compress trim tokens, keep instructions
  3. 3 · Classify complexity 0 / 1 / 2
  4. 4 · Route cheapest model that can do the job
  5. 5 · Batch & log group requests, record cost saved
Full stack · Applied AI

ReachScout

A professional-research assistant that finds the right recruiter, hiring manager or founder from public sources, shows the evidence behind each data point, and drafts a context-aware introduction. Built to be explicitly not a scraper or spam tool.

  • Five-stage search pipeline: normalize, generate queries, retrieve, extract, rank and dedupe.
  • Evidence-backed confidence scores; estimated emails are never labeled verified.
  • Security first: SSRF blocklist, row-level security, per-user rate limits.
TypeScriptNext.js 15SupabasePrismaClaude APIVitestPlaywright
Tested guarantees
  • Duplicate profiles merge
  • Pattern-guessed emails never marked verified
  • Private-network URLs rejected
  • Dangerous search operators stripped

Everything on GitHub

Pulled live from github.com/dmeltonyan.

Experience

Full resume →

Jun – Oct 2025

Machine Learning Intern · Builders and Tradesmen's Insurance Services

  • Shipped a Python classification service mapping plain-English business descriptions to 691+ WCIRB class codes, replacing a fully manual underwriting step.
  • Built confidence-based routing: certain results auto-resolve, ambiguous ones go to a human. Cut manual review time about 40% with no silent misclassification.
  • Designed and tuned PostgreSQL schemas for model outputs; automated ETL processing 20K+ records a month.
Mar – Jun 2025

ML / GTM Intern · Retail Labs

  • Built REST endpoints and SQL aggregation layers feeding internal dashboards, roughly halving analyst turnaround on recurring reports.
  • Built document-parsing pipelines turning unstructured quarterly reports into normalized JSON for BI tools.
  • Contributed to backend ingestion for the company's AI retail agent.
Nov 2024 – Mar 2025

Automation / Test Engineer Intern · Pacific Coast Companies

  • Built Python tooling to generate production-like synthetic data for regression tests, so no customer data reaches lower environments.
  • Co-led automated test coverage for critical SAP SuccessFactors workflows.
Jun 2023 – Nov 2024

Coding Instructor · Code Ninjas

  • Taught Python, JavaScript and Scratch to 100+ students aged 7 to 14; built the instructors' progress-tracking dashboard.

Toolkit

Languages

PythonTypeScriptJavaScriptJavaSQL

ML & data

scikit-learnLightGBMPyTorchPandasSHAPBERT embeddings

Backend

FastAPIFlaskSpring BootExpressRedisREST

Data & infra

PostgreSQLSupabaseDuckDBDockerCI/CDNext.js

Hiring for ML or backend in 2027? Let's talk.