On-device speech AIActive
Ayah · آية
An offline-first “Shazam for the Quran” that identifies the exact Surah and Ayah from recitation using a Whisper-class model running entirely on-device.
React 19Transformers.jsONNXPWAOffline-first
AI Engineer · Founder, Nexasity AI
Daud Ibrahim HassanAI Engineer — LLM evaluation, calibration & reliability
Production experience across RAG pipelines, multi-LLM orchestration, multi-agent systems, and on-device inference. Shipped 10+ AI/ML systems to 500+ users; code merged into Grafana and Crawl4AI. BSc Computer Science, BRAC University.
Why me
I build AI for the people the conversation usually leaves out — on-device inference that never leaves your phone, retrieval that reaches learners priced out of coaching, and systems that verify instead of hallucinate.
On-device & private
Inference that never leaves the phone. Ayah runs a full-precision ONNX model on-device via Transformers.js — audio is transcribed locally and never uploaded.
Reaches underserved learners
RAG for students priced out of coaching. LesLearn and RootLearn ingest the national curriculum and teach from each learner's actual level — offline and zero-cost by design.
Verifies, not hallucinates
Evals, calibration, and grounded retrieval. whyLayer, CiteXai, and the World Cup 2026 calibration loop check answers before trusting them.
Four systems, up close
World Cup 2026 → whyLayer → Ayah → skeletree
- 01World Cup 2026 — calibrated match predictor
- 02whyLayer — verified decision engine
- 03Ayah — on-device Quran recognition
- 04skeletree — token-cheap repo maps for agents
Ayah captures ~7s of audio and runs a ~145MB ONNX model on-device via Transformers.js. Arabic recitation is transcribed and fuzzy-matched to the exact Surah and Ayah — audio never leaves the phone.whyLayer interrogates a decision through 3–9 adaptive questions, hunts web evidence (Serper / Jina / DuckDuckGo), and returns a GO / NO-GO verdict with a confidence score — reasoning routed across a 5-provider LLM cascade.
World Cup 2026 predicts every match with a 5-model stacking ensemble over 40+ engineered features — evaluated on log-loss, RPS, Brier, and ECE, with post-hoc calibration (temperature + isotonic) and a daily retrain-and-monitor loop.
skeletree scans a repo once and emits a compact map — the directory tree plus every class and function signature, no bodies — so AI coding agents grasp a codebase for ~90% fewer tokens. Published on PyPI with MCP integration.What I build
Four themes run through the work: private on-device AI, decision intelligence you can trust, calibrated quant-ML, and developer tooling.
On-device AI
Private inference that runs entirely on the phone — audio and data never leave the device. Ayah, RootLearn, RainXai.
Decision intelligence
Multi-LLM systems that interrogate and verify before they answer. whyLayer, False9.
Quant & applied ML
Calibrated models with proper scoring, time-split validation, and monitoring. WC2026, StockIND, HR Attrition.
Developer tooling
Tools that make AI agents cheaper and sharper. skeletree, oss-hunter, GitChrono.
All projects
Everything I’ve shipped or am building — 32 projects across on-device AI, decision intelligence, quant-ML, RAG, and developer tooling.
Decision-intelligence engineActive
whyLayer
A decision engine that interrogates a choice before you make it — multi-LLM profiling plus web research, returning a GO / NO-GO verdict with a confidence score.
React 19FastAPIMulti-LLMGroqGemini
Quant research terminalActive
StockIND
An AI stock scanner that screens 500+ symbols with per-ticker XGBoost models and a FinBERT sentiment veto. Research only.
PythonXGBoostFinBERTFlaskGroq
Multi-agent code auditorActive
False9
An autonomous system that audits, fixes, and completes projects with 12 domain-expert agents running in parallel inside any AI coding environment.
Multi-agentAI-codingPrompt-engineeringMCP
Prompt-engineering platformActive
codePromptx
A gamified platform for learning prompt engineering across 100 challenges with 3-layer AI grading (validation → regex → LLM judge), XP, and certificates.
React 19SupabaseGroqExpress
Repo-mapping CLI for agentsActive
skeletree
A CLI that maps any repo into a token-cheap skeleton — tree plus every signature, no bodies — cutting AI-agent context cost ~90%. Published on PyPI.
PythonPyPItree-sitterCLI
AI citation verificationWIP
CiteXai
Detects fake and AI-hallucinated citations by cross-referencing them against 200M+ papers across CrossRef, OpenAlex, Google Books, and arXiv.
Next.js 16React 19GroqRAG
Sports-analytics MLActive
World Cup 2026 Predictor
A 5-model stacking ensemble over 40+ features with strict time-split validation and post-hoc calibration — evaluated on log-loss, RPS, Brier, and ECE.
XGBoostLightGBMKerasEnsemble
Offline Bangla math tutorWIP
RootLearn
A Bangla-first adaptive math tutor (Class 5–8) that diagnoses learning gaps and teaches from the student’s actual level — fully offline via local Gemma.
React 19FastAPIOllamaRAG
RAG learning platformWIP
LesLearn
A RAG adaptive-learning platform for Bangladeshi students (Classes 1–10) priced out of coaching — zero-cost by design.
RAGChromaDBNext.jsEdTech
Document intelligence / RAGWIP
Legal Doc Intelligence
Ingests messy legal documents, retrieves evidence via vector search, and generates grounded case summaries with a citation per claim.
PythonChromaDBOCRRAG
Ticket classifier APIActive
QueueStorm
A deterministic, rules-based CRM ticket classifier — problem type, severity, department, plus phishing detection in English and Bangla. No LLM.
FastAPIPythonDockerRules-engine
OSS contribution CLIActive
oss-hunter
A CLI that finds real good-first-issues, ranks them by ease of merging, and drives the workflow from discovery to pull request.
PythonCLIGitHubOpen-source
Android call analyticsActive
Calyz
An Android app that turns raw call history into ranked leaderboards and communication analytics — 100% on-device.
KotlinJetpack ComposeRoomMVVM
Android screen auditorWIP
Sapphire
An Android screen-content auditor and app-locker that uses Gemini to label what’s on screen Useful or Wasteful and lock out lazy scrolling.
KotlinJetpack ComposeGeminiAndroid
AI browser extensionActive
NexMark AI
A Chrome extension that auto-categorizes bookmarks with a three-tier engine of domain matching, regex rules, and Groq LLaMA 70B.
Chrome ExtensionGroqJavaScriptMV3
Developer productivity CLIActive
GitChrono
Estimates coding time from GitHub language stats. Published as a CLI on npm with a web dashboard.
TypeScriptNext.jsPrismaCLI
PDF → video pipelineWIP
YouTube Automation
An end-to-end pipeline converting PDF content into narrated, animated videos via CrewAI → Manim → EdgeTTS, with auto-upload.
PythonCrewAIManimSelenium
EdTech marketplaceActive
WhatTheProject
An academic marketplace connecting Bangladesh students with vetted experts, with AI price estimation and bKash payments.
Next.jsSupabaseGroqbKash
Anonymous messagingWIP
Unkahi
An anonymous messaging platform tied to Instagram handles — no signup, private bookmark-only inbox.
Next.jsSupabaseFramer MotionTS
Jira predictive analyticsWIP
dejaView
A Jira app that analyzes project data and generates predictive “future retrospectives” so teams anticipate risk early.
Atlassian ForgeReactJiraConfluence
Lead-gen automationWIP
Reddit Intent Engine
Scans Reddit for buying-intent posts, scores them, and drafts compliant outreach — draft-only by default.
PythonBeautifulSoupSQLiteAutomation
D2C skincare storefrontActive
Lunaria
A direct-to-consumer storefront with geo-adaptive pricing and an admin panel, shipped as a single-file build.
ReactSupabaseViteTailwind
Personal finance web appWIP
SpendXai
A client-side expense tracker with a clean dashboard, multi-currency support, and a Zod-validated entry form.
ReactTypeScriptshadcn/uiRecharts
Voice expense trackerWIP
RainXai
A privacy-first expense tracker where you log spending by voice in English or Bengali, stored entirely on device.
KotlinJetpack ComposeVoiceOffline-first
Supervisor matchingWIP
ThesisMap
A supervisor-student matching engine that pairs thesis supervisors with students by research-interest overlap.
ReactExpressSQLiteTypeScript
Education dashboardArchived
StudentProgress · The Climb
An offline-first dashboard that ranks students by mean score and exports an LLM-ready evaluation prompt in one click.
ReactFramer MotionViteTS
ML classificationArchived
HR Attrition
A CRISP-DM attrition model on the 1,470-record IBM set — XGBoost at 92% accuracy / 0.92 ROC-AUC with a Tableau dashboard.
PythonXGBoostSMOTETableau
ML notebookArchived
CO₂ Regression Lab
A linear-regression notebook predicting vehicle CO₂ emissions from engine size, cylinders, and fuel consumption.
Pythonscikit-learnJupyterRegression
Law-enforcement systemArchived
CAIS
A web app for tracking criminal profiles, investigations, and evidence, and finding crime patterns — built on raw SQL.
DjangoRaw SQLMySQLTailwind
Website templateWIP
Agricko Template
An agriculture and organic-farming marketplace PHP template with transporter and delivery-management pages.
PHPBootstrapMySQLSCSS
Location-time trackerWIP
wandR
An Android app that tracks where you go using free OSS APIs and shows where your time went — raw GPS never leaves the device.
React NativeExpoFirebaseTS
The toolkit
The stack behind the work, and the loop it runs on: Data → Model → Eval → Deploy → Monitor.
LLM Evaluation & ReliabilityProbability CalibrationRAG PipelinesOn-device InferencePython
scikit-learnXGBoostLightGBMTensorFlow / KerasFinBERTsentence-transformersONNXTransformers.jsNumPyPandasFeature engineeringHyperparameter tuningCross-validationSMOTEPCAStacking ensemblesTime-split validationTemperature / isotonicECELog-loss / RPS / BrierModel monitoringDrift & retraining loopsVersioned artifactsChromaDBSemantic chunkingGrounded generationMulti-LLM cascade routingGroqGeminiSambaNovaOpenRouterOllamaCrewAILangChainPrompt engineeringArabic speech recognitionCitation verificationSentiment analysisMisconception detectionReact 19Next.jsTypeScriptTailwindViteFastAPINode.jsExpressRESTPHPSQLPostgreSQLSQLiteTableauPower BIMatplotlibSeabornPlotlySupabaseFirebaseDockerVercelGitGitHub ActionsKotlinJetpack ComposeAndroid SDKCapacitorEnglish (fluent)Bengali (native)Urdu (conversational)
Certifications
Credentials in agent tooling, applied AI, and data analytics — from Anthropic, Google, and industry programs.
AnthropicJun 2026
Introduction to Model Context Protocol
MCP fundamentals — exposing tools & resources and wiring agents to live data.
Anthropic · CodePathJun 2026
AI Fluency for Builders
Building effectively with AI — prompting, agent workflows, and responsible use.
Google · CourseraJul 2026
Google AI Essentials
5-course specialization — AI tools, prompting, and responsible AI use.
Analyst BuilderAug 2025
Data Analytics Bootcamp
SQL, Excel, Tableau, Power BI, and exploratory data analysis.
Kind words
(coming soon)
“Testimonial coming soon — this space is reserved for feedback from collaborators and the people who use what I build.”
