Work

Built, measured, and stress-tested

A production knowledge platform still leads the work. Around it are independent projects in model training, evaluation, retrieval, and AI infrastructure, each scoped by evidence rather than a demo claim.

Flagship · Search & retrieval platform01

Shia LibraryLive

Web · iOS · Android

A production digital library that grew from search into answers — and from one client into two. Diacritic-insensitive Arabic search, English full-text, and semantic vector search over a corpus of over 315,000 ingested records; a grounded "Ask" (in beta) that retrieves from the live corpus and answers with Claude's native citations, so every claim links to the exact source span or the system abstains; and one Supabase/Postgres backend serving both the web platform and offline-first iOS and Android apps. Built with row-level access control, an in-app CMS, and a reading experience tuned for dense scholarly text.

Next.js 16React 19TypeScriptSupabase / PostgrespgvectorClaude (native citations)

~315k: hadith records ingested
91%: claims grounded (30-q eval)
1,000+: web tests in CI
0.94: faithfulness (Claude judge)

Read case study

Machine learning & evaluation02

PolaritySource release

A 184M-parameter cache-equivalence model, and the scientific debugging trail that established both where it works and where it fails.

PythonPyTorchHugging Face TransformersDeBERTa-v3scikit-learn

184M: trainable parameters
2/44: Veritas false accepts

Case study

Mobile engineering03

Shia Library — Mobile AppLive

iOS · Android

The same library, offline-first on iOS and Android — hand-written Swift and Kotlin keep downloads alive under each OS's rules, and the offline store makes a half-visible book impossible. Live on the App Store and Google Play.

React Native (Expo)TypeScriptSwiftKotlinObjective-C++

3,900+: automated tests (Vitest)
iOS + Android: public, one codebase

Case study

AI / retrieval systems04

Citation-Grounded RAGLive

A retrieval-augmented answer engine where every claim is grounded in a specific source — hybrid retrieval, verifiable citations, an abstain path, all measured by an eval harness.

Next.js 16TypeScriptClaude (native citations)OpenAI embeddingsBM25 + RRF

0.91: recall@5 (hybrid)
1.00: faithfulness (Claude, 22-q)

Case study

AI infrastructure05

LLM GatewayLive

A provider-agnostic LLM gateway whose semantic cache is proven correct — precision, false-positive rate, and a CI gate most managed gateways don't expose.

Next.js 16TypeScriptClaude Haiku (intent judge)OpenAI embeddingsOpenTelemetry gen_ai.*

1.00: cache precision (guarded)
0%: false-positive rate

Case study

AI evaluation & methodology06

JudgelabLive

A reproducible lab that measures how reliable LLM-as-a-judge systems actually are — chance-corrected agreement with confidence intervals and a keyless, license-clean benchmark, not a single agreement score.

Pythonuv · Ruff · mypy (strict)pytestNumPy / SciPyPydantic · Typer

1,814: aligned judge–human comparisons
κ 0.767: GPT-4 vs human (ties excluded)

Case study

Data engineering & LLM orchestration07

Usul PipelineLive

A resilient, cost-optimized ingestion + LLM-translation pipeline that turns scattered source texts into a clean, structured corpus.

Pythonasyncio / httpxClaude (Anthropic)Supabase / PostgresPrompt caching

220: books processed
35k+: passages translated

Case study

On the roadmap

What I'm building next

These aren't shipped yet — listed honestly. The first items deepen the platform; alongside them I'm branching into independent tools that solve broader problems. Depth and range, not scattered demos.

Finish the corpus: work the book-by-book validation queue that gates publication, and re-ingest + embed the excluded large book so semantic recall covers the whole library.
If weight release becomes worthwhile, retrain from a provenance-clean data recipe under a fully captured environment and review the resulting checkpoint rights separately from the code license.
Ship daily-hadith push notifications (designed, not yet built) and grow the Maestro end-to-end suite in CI, now that both public store listings are live.
Deep-link each citation chip straight to its source page in the reader, now that the production corpus path is live.
Push the learned-threshold work further: the leave-one-out experiment shows cheap features cap at recall 0.32, so the path is a vCache-style per-prompt boundary on a richer labelled set — can it close the gap to the judge without the judge call?
Wire the owner-gated live-judge path (OpenAI / Anthropic adapters behind cost caps) and run the bias battery on fresh orderings: position, verbosity, self-preference, and test-retest probes.
Extract the ingestion core into a config-driven, source-agnostic open-source framework with a typed SDK and CLI.
More broadly — distributed-systems and real-user products in new domains, to build range beyond a single platform.

See experiments →