Hamna Raeel builds
intelligent systems with
deep learning & data.

Hirschsprung Disease Detection via Segmentation‑Guided Histopathology Analysis

A Mask2Former-based segmentation model with a ResNet–Swin Transformer backbone, paired with a downstream classification framework for automated histopathology disease detection.

Try live inference on Hugging Face

Screenshots above show real output — live Space may take ~30s to wake from idle.

View Source on GitHub

Meta‑Learning for Zero‑Shot Cross‑Lingual Transfer (CML‑mDeBERTa)

A Curriculum Meta-Learning framework built on first-order MAML, trained on the XNLI dataset to enable zero-shot natural language inference across 15 languages.

Try live inference on Hugging Face

Live Space may take ~30s to wake from idle.

View Source on GitHub

Sentence Classification using Convolutional Neural Networks (TextCNN)

A TextCNN model for sentence classification on the AGNews dataset, benchmarked across random, static, non-static, and multichannel word embedding strategies.

Link to be added

Prompt Versioning & A/B Testing Platform

Most teams tweak an LLM prompt, eyeball a few outputs, and ship — with no way to know if the change actually helped, quietly hurt accuracy for a slice of users, or tripled the token bill, until someone complains. This platform brings feature‑flag‑style rigor to prompt changes instead: every version is saved like a Git commit, traffic is split between variants with consistent hashing, and a real stats engine (Welch's t‑test + Mann‑Whitney U) declares a significant winner before it auto‑promotes to production.

View live dashboard View Source on GitHub

LLM Cost Autopilot — Complexity‑Routed LLM Gateway

Every request to a large model like GPT‑4 costs the same whether it's answering "what's 2+2" or a genuinely hard question — most production traffic doesn't need the expensive model, but routing prompts to cheaper ones by hand is risky with no safety net. This gateway solves that: it scores each prompt's complexity with an interpretable RandomForest classifier, sends it to the cheapest model that can handle it, and runs an async verifier that escalates to a stronger model whenever the cheap model's answer looks wrong.

View live dashboard View Source on GitHub

LoRA Fine‑Tuning Pipeline for Automated Code‑Review Revisions

Base LLMs asked to apply a reviewer's comment don't just fix what was flagged — they reformat neighboring lines, add unrequested error handling, and wrap the diff in prose. This pipeline QLoRA‑fine‑tunes Mistral‑7B on 2,500 high‑signal examples curated from 150k+ raw PR diffs via a custom grounding‑score filter, guarded by three independent data‑leakage checks, so the model reproduces untouched code byte‑for‑byte and changes only what was asked. Evaluated with paired bootstrap confidence intervals and McNemar tests against a benchmark built so a copy‑the‑input baseline can't score well.

View Source on GitHub

LLM API Gateway — Rate Limiting, Failover & Observability

Most teams call OpenAI or Anthropic directly from application code, then discover the hard way that they need rate limiting, budget guardrails, and a failover plan for when a provider degrades. This gateway is that infrastructure layer: a unified proxy across OpenAI, Anthropic, and Ollama with Redis‑backed atomic per‑team rate limiting, budget caps and per‑model access control enforced before a request reaches a provider, and automatic fallback routing verified end‑to‑end during a simulated provider outage. Circuit breakers run the full closed → open → half‑open → closed lifecycle, with every request, cost, and failover event traced via OpenTelemetry and surfaced across three Grafana dashboards.

View Source on GitHub