Résumé — Abdi Alamir

Summary

AI Engineer who builds and evaluates production Large Language Model (LLM) systems end to end — retrieval pipelines, tool-calling agents, and the backends they run on. Particular focus on finding where a model or system is confidently wrong: calibration errors, silent retrieval failures, and ungrounded generation. Ships to production on Microsoft Azure with typed schemas, test coverage, and CI/CD. Master’s in Data Science. Four years running my own licensed service business before engineering, which is why I build for the customer’s actual workflow rather than the demo.

Technical Skills

Languages Python, TypeScript, SQL, JavaScript, HTML, CSS

LLM & AI LangGraph, LangChain, Retrieval-Augmented Generation (RAG), Model Context Protocol (MCP), pgvector with HNSW indexing, vector embeddings, tool-calling agents, agentic workflows, prompt design, structured output, conversational AI, Anthropic Claude API, OpenAI API, Azure OpenAI Service, LightGBM, SHAP

Evaluation Golden datasets, held-out evaluation sets, backtesting harnesses, calibration analysis, grounding and hallucination control, provenance tracking, retrieval metrics (recall@k, precision@k, MRR)

Backend & Data FastAPI, PostgreSQL, SQLAlchemy, Alembic, REST API design, async Python, pytest, Redis, web scraping, structured data extraction, dimensional modeling

Infrastructure Microsoft Azure (App Service, Container Apps, Static Web Apps, PostgreSQL Flexible Server, Key Vault), Docker, GitHub Actions, CI/CD, Prometheus-format metrics, observability

Frontend Next.js, React, Tailwind CSS, TypeScript

Experience

AI Engineer

Alamir Labs — independent practice · Remote

2024 — Present

Design, build, and deploy production AI systems end to end, from the data layer through model orchestration to shipped user interface.

Prediction-Market Signal Engine

Python · FastAPI · PostgreSQL · Next.js · Microsoft Azure

Built a market-scanning engine evaluating 720+ live Kalshi markets for constraint arbitrage and distributional mispricing, with 1,391 passing tests under CI/CD on Azure.
Diagnosed a calibration failure in my own model: the strategy showed a 64% win rate while losing money. Root cause was a units error in the edge definition — the system detected a basket arbitrage but executed a single leg, so the basket’s edge was being quoted on a directional trade. True expected value was −2.9% after fees, and realized P&L matched that prediction rather than the win rate.
Implemented a verifier rule at the order choke point to reject negative expected-value legs, and reported the negative result honestly rather than shipping the win-rate headline.
Built a hybrid retrieval pipeline (embeddings + BM25, score fusion, reranking). Recall@6 improved from 0.47 to 0.85 on a golden set, and 0.50 to 0.75 on a held-out paraphrase set written specifically to detect overfitting to my own design choices.
Exposed the governed toolset over a Model Context Protocol (MCP) server so other AI clients query the same certified metrics instead of re-implementing them.

Cross-Data Concierge Agent

LangGraph · pgvector · Azure OpenAI · FastAPI — deployed to production

Built a LangGraph tool-calling agent that answers questions spanning unstructured documents and live business data in a single turn, routing between a pgvector/HNSW RAG retriever and a SQL analytics tool without the user specifying a source.
Shipped to production on Azure App Service and PostgreSQL Flexible Server. The endpoint returns the answer plus tool trace and source citations, so every response is auditable.
Found and fixed a silent retrieval failure: a cosine-similarity cutoff of 0.70, tuned for a prior embedding model, filtered out every correct match after a model swap because the new embeddings scored real matches at 0.35–0.65. Retrieval returned nothing and the agent answered from parametric memory instead of the knowledge base.
Built a governed semantic layer with one certified definition per metric, so the same question returns the same correct answer every time.

Grounded Business-Assistant Generator

Python · FastAPI · Anthropic Claude API · structured extraction

Built a pipeline that ingests a business’s public website and produces a working, grounded chat assistant: scrape → typed profile extraction → render → served chat backend.
Measured the deterministic baseline before adding an LLM across 12 real business sites: 10/12 parsed from a plain fetch, 9/12 published typed schema.org markup, yielding 68% average profile completeness with zero LLM calls. Used the measurement to scope the model to gap-filling only.
Enforced per-field provenance so generated content can never overwrite verified source data, and designed the assistant to decline questions it cannot ground rather than fabricate an answer — verified end to end against adversarial questions.

Platform & Production Engineering

Trained and evaluated a LightGBM classifier with point-in-time feature engineering, walk-forward validation, and exact TreeSHAP explanations — then recommended against deployment, with measurements showing it underperformed a simpler baseline at the available data volume.
Added observability — counters, gauges, latency histograms and separate failure counters in Prometheus exposition format — after two silent faults went undetected for weeks behind logs alone.
Audited the system against live production data and identified four failure modes: a detector misreading market structure, a silent grading no-op, a schema overflow rolling back every write, and a model running on a hardcoded constant 3.5× off the measured value. Each was reproduced with data, fixed or retired, and covered by regression tests.
Migrated the production database across providers with no application code change.

Owner and Operator

Abdi Massage · Portland, OR

2015 — 2019

Founded and ran a State of Oregon licensed massage therapy practice for four years, employing two staff.

Operated as a fully licensed practice under the Oregon Board of Massage Therapists. Held an active Oregon LMT license for the life of the business and hired only licensed therapists, in an industry where unlicensed operation is common.
Owned all business operations: hiring, staff scheduling, pricing, billing, client intake, records, and state compliance.
Built and retained the client base directly, managing repeat bookings and long-term customer relationships.
Direct experience running the kind of appointment-driven service business that later informed my work building AI front-desk and booking assistants.

Education

Master of Science, Data Science for Business

Portland State University · Portland, OR

2023 — 2025

Bachelor of Science, Business Technology and Analytics

Portland State University · Portland, OR

2019 — 2023