Noah Dericioglu Portfolio

Software engineer building AI applications, inference systems, and evaluation tools.

Irvine, California

Projects

01 / 06

Six projects, one active card at a time. Select a title, then turn the card to read more.

Project index

Read project details
A documentation answer should let the reader inspect its source. This application connects a React/TypeScript interface to a Python/FastAPI backend and Weaviate retrieval, returning generated answers with source citations.
The retrieval evaluation compared hybrid and vector search on documentation questions. It measured retrieval accuracy and p95 query latency under concurrent requests, then examined search failures to identify improvements.
The app was containerized with Docker and its Kubernetes deployment was tested. Integration tests cover retrieval failures and model API timeouts. A runnable tutorial was published and the findings were presented to the engineering team; the tutorial is not linked in this collection.
Limit of claim: Kubernetes deployment was tested. No production cluster scale or numerical retrieval and latency results are published.
Read project details
Reviewing ship-bridge safety events means finding relevant moments inside voyage recordings. This system connects an AWS ingestion pipeline to a web portal where operators can inspect detected incidents.
C++ frame preprocessing feeds a detection model served through NVIDIA Triton Inference Server. Predictions become timestamped video events, connecting model output to footage an operator can review. The build covered data collection, model training, validation, and production deployment.
The evaluation benchmarked inference latency and throughput across batch sizes, and examined detection failures using footage from the Istanbul-Trieste route. Hardware details and numerical benchmark results are not part of this public case study.
Limit of claim: Latency and throughput were benchmarked across batch sizes. Numerical results, hardware configuration, and fleet-wide deployment scale are not published.
Read project details
Can model-generated strategy proposals add value once risk, costs, and evaluation are controlled? This Python paper-trading research system keeps the LLM agents that propose and critique strategies separate from the deterministic code that enforces risk limits and evaluates performance.
A separate GKE serving demonstrator was verified on October 9, 2026: Qwen2.5-1.5B-Instruct served by vLLM 0.31.0 on one NVIDIA L4. Kubernetes requests the GPU, checks model readiness, and mounts a 20 GiB persistent model cache. Recorded CUDA execution and model generation verify the runtime; replacing the model pod preserved the cache marker and restored serving.
An authenticated FastAPI gateway validates fixed task contracts, limits prompt and output sizes, checks the backend token count, and enforces a 30-second deadline with disconnect cancellation. It admits eight requests at a time and returns HTTP 429 for excess requests. The public repository includes the gateway, manifests, runbook, benchmark driver, and runtime records.
The concurrent benchmark retains 128 measured attempts and excludes four warmups. At concurrency eight, 32 of 32 requests succeeded at 10.864 successful responses per second, with a successful-response p95 of 791 ms. At concurrency sixteen, eight succeeded and 24 returned HTTP 429. This was one repeated short sentiment prompt, not a varied financial-news corpus; client timings include the Kubernetes port-forward, network path, and validation.
Five paper strategy books share a universe: a rules-based control, news sentiment, LLM proposals, post-filing event research, and a combined book. Anthropic/Ollama and deterministic clients remain available; vLLM is an opt-in provider. A separate proposer/critic/portfolio demonstration used explicitly synthetic inputs and the deterministic risk gate, without submitting orders or measuring investment returns. Its output passed the JSON contracts but missed semantic instructions: four rules instead of three, a critic that largely repeated them, and two positive portfolio weights when opposite signs were requested. Those failures identify the next evaluation work.
The evaluation policy requires at least 60 trading days and p below 0.05 on a paired Diebold-Mariano test before a superiority claim. That threshold alone does not establish fair risk-adjusted comparisons or a demonstrated edge; the research history includes corrections to earlier results.
In the public Python research code, the risk wrapper rejects non-finite weights and symbols outside the briefing universe, clips position and liquidity exposures, scales gross exposure, and throttles turnover against the current book. The opt-in vLLM adapters record successful model calls separately from deterministic heuristic or lexicon fallbacks. These controls constrain proposals; their existence alone does not establish a strategy edge.
Limit of claim: The serving benchmark uses one repeated short prompt in a brief isolated run. It establishes neither production capacity nor trading performance. The demo was cleaned up after verification; the paper-book history is not a record of continuous vLLM service.
View repository
Serving code
Benchmark & scope
GPU runtime record
Deployment runbook
Read project details
When does an earnings-call answer evade the question? This classifier compares six LLMs on zero-shot, three-class evasiveness detection using EvasionBench: GPT-4o, GPT-4.1, Claude, Gemini, DeepSeek, and Grok. A fine-tuned RoBERTa-large model provides a supervised baseline.
A stacking ensemble combines the LLM predictions and RoBERTa features. It achieved 0.81 F1, above every individual model in the project evaluation. Prompt refinements addressed ambiguity between categories and classification consistency.
The project was advised by Prof. Daniel Ringel. F1 averaging, split design, and individual model scores are not included in the public summary.
Limit of claim: A capstone result, not a production deployment. The published summary does not specify F1 averaging, split design, or individual model scores.
Read project details
Pari turns a wine-label photo into structured wine information and builds recommendations from a user’s tastes. The SwiftUI app sends label analysis through Supabase Edge Functions to Claude, keeping model credentials on the server.
PostgreSQL and pgvector support similarity-based recommendations, with a fallback when a user has limited rating history. Wine embeddings use 64 dimensions, combining varietal and region semantics with observed tasting structure; hashing covers unknown terms.
Taste matching combines shared-rating correlation with content similarity. The repository includes the iOS app, Edge Functions, database migrations, and evaluation scaffolding. It is source-available for inspection, not open-source licensed. Whether this personalized rating predicts taste better than the crowd average still needs more observations.
In the current public source, label and wine-list requests require a validated user JWT, pass an image-size check, and consume a shared per-user quota before calling the provider. Model output is parsed and checked against explicit object and field contracts; malformed responses return a controlled error. A 30-second deadline covers the provider request and response body. These are current source-code behaviors, not a claim that the latest changes have been independently tested in the deployed app.
Limit of claim: That the twin-weighted rating beats the crowd average. Too few bottles have been logged to test whether leaning on the users whose palate tracks yours predicts taste any better.
View repository
Read project details
Maritime route planning and fuel monitoring depend on turning vessel measurements into useful operating information. This project combines predictive models for route optimization with sensor analysis for fuel-efficiency planning and CO2 monitoring.
Shipping-lane and container-flow dashboards bring the analysis into recurring operations reviews. Numerical fuel savings, emissions reductions, and model scores are not published in this case study.
Limit of claim: The work supported fuel-efficiency planning and CO2 monitoring. No numerical fuel savings, emissions reductions, or model performance are published.

Ask the desk

Selected workSelect a work to focus the desk.

Explore the architecture, decisions, and results.

Ask how a project works, why it was built that way, or what its evaluation shows.

Enter to send · Shift + Enter for a new line

Automated assistant · Chat history clears on reload · email Noah

Notes

The work is technical, but the through-line is judgment: find the uncertain part, make it legible, and test it before making the claim.

These are working systems and honest records of their limits. Some have shipped into operations, some are still in development, and each entry keeps that distinction visible.