SELECTED WORK · CASE STUDIES

The Problems
We Solved.

Six projects that required domain expertise nobody else had. Each one started with a problem that the standard vendors couldn't touch.

01 AIDA — Document Intelligence Platform Oil & Gas · AI/ML 02 Predictive Maintenance ML — TAQA Well Solutions Energy · Operations 03 SAP Digital Transformation — AlMansoori Oil & Gas · ERP 04 Cybersecurity Audit — BLS Events / First Arab Bank Finance · Security 05 Planetary Ephemeris Engine — High-Precision Astronomy Scientific Computing 06 AISHA — Intelligent Web Automation Automation · Systems
AIDA
Vision-Language · Knowledge Graphs · Rust/Go
CASE STUDY / 01

AIDA — Document
Intelligence Platform

60,000 oilfield documents. Decades of technical manuals, well reports, inspection logs, compliance certificates. All locked in formats that no search engine could read and no LLM had ever seen. The mission: make them queryable in natural language with 95%+ extraction accuracy.

Industry
Oil & Gas
Domain
Document AI
Scale
60k Documents
Accuracy
95%+
95%
Extraction Accuracy
60k
Documents Indexed
<2s
Query Response Time
The Problem

An oilfield service company had 60,000+ documents accumulated over decades — technical manuals, well completion reports, HSE records, inspection certificates, equipment logs. Institutional knowledge was trapped inside PDFs and scanned images. Every time an engineer needed historical data, they searched manually, taking hours or days.

Existing document management systems could store files but couldn't understand them. Commercial OCR tools missed domain-specific terminology. Generic LLMs hallucinated when asked about oilfield-specific data formats.

The Approach

Built a purpose-trained vision-language model pipeline using Gemma3, fine-tuned on oilfield document vocabulary. Custom OCR preprocessing handles degraded scans, rotated pages, mixed Arabic/English content, and technical diagram annotations — problems that broke every commercial tool tested.

Structured the extracted knowledge into a property graph (SurrealDB) linking equipment IDs, well identifiers, inspection dates, personnel, and certification chains. Natural-language queries resolve against both semantic embeddings and graph traversals simultaneously.

The Outcome

95% extraction accuracy on domain-specific content — verified against ground-truth annotations by subject matter experts. Query latency under 2 seconds for full corpus search. Engineers who spent hours searching now get answers in seconds with source citations.

The knowledge graph surface also revealed compliance gaps: equipment without current certifications, overdue inspection schedules, incomplete traceability chains — none of which were visible in the raw document pile.

Stack Rust Go SurrealDB Gemma3 Vision-Language Models Knowledge Graphs Custom OCR Semantic Embeddings Arabic/English NLP
ML
Predictive Maintenance · SAP Integration · Python
CASE STUDY / 02

Predictive Maintenance
ML — TAQA Well Solutions

Downhole tool failures in active well operations mean lost production, workover costs, and safety risk. The goal was to predict failures before they happened — and integrate that intelligence directly into the SAP plant maintenance workflow engineers already used.

Client
TAQA Well Solutions
Domain
Predictive ML
Integration
SAP PM
Result
35% Downtime ↓
35%
Downtime Reduction
72h
Avg Failure Lead Time
SAP
PM / AM / MM Integrated
The Problem

Downhole tools — perforating guns, logging tools, packers — fail under extreme pressure and temperature. Maintenance was reactive: run a tool until it failed or returned from a job showing wear. Failure prediction relied entirely on engineer intuition and calendar-based schedules regardless of actual usage intensity.

SAP PM contained years of maintenance history, but nobody had ever built a model over it. The data existed; the insight didn't.

The Approach

Extracted structured maintenance history from SAP PM going back five years — 12,000+ maintenance orders, tool run records, failure codes, repair actions. Built Python data pipelines to clean, normalize, and feature-engineer the raw SAP data into model-ready form: run counts, cumulative pressure-hours, failure-to-repair gaps, tool class clusters.

Trained gradient-boosted classifiers on failure precursors. The model outputs a risk score per tool per week, surfaced through a .NET dashboard that integrates directly with SAP to auto-generate preventive maintenance orders at threshold risk levels.

The Outcome

35% reduction in unplanned downtime within the first operational quarter. The model reliably flagged high-risk tools 48–72 hours before failure events confirmed in the historical test set. Maintenance teams shifted from reactive repair to planned intervention.

The SAP integration meant zero workflow change for field engineers — they received standard PM work orders in the system they already worked in. Adoption was immediate because nothing changed except the source of the work order.

Stack Python SAP PM/AM/MM Gradient Boosting C#/.NET SQL Server SAP RFC Feature Engineering Time-Series ML
SAP
SAP AM · MM · PM · Ariba · Logistics
CASE STUDY / 03

SAP Digital Transformation
— AlMansoori Specialized Engineering

A regional oilfield services company with 5,000+ assets, a global supply chain, and maintenance operations across six countries. The ERP was SAP — but SAP alone was not enough. The gap between the system of record and operational reality needed closing.

Client
AlMansoori Group
Duration
6 Years
SAP Modules
AM / MM / PM / Ariba
Scope
5,000+ Assets
5k+
Assets Under Management
6
Countries Covered
100%
Asset Traceability
The Problem

Large oilfield services companies live and die by asset availability. At AlMansoori, the challenge was three-fold: assets moved across borders without proper system updates, procurement and maintenance were running on disconnected workflows, and CapEx planning was based on spreadsheets that nobody trusted.

SAP was nominally in place, but utilization was shallow. Critical modules — AM for asset lifecycle, Ariba for procurement, Logistics for cross-border movement — were either underused or bypassed entirely by field teams who'd found workarounds.

The Approach

Led the standardization of SAP Asset Management processes across all business units. Built custom .NET applications to bridge SAP workflows with field operations: barcode and RFID scanning for real-time asset check-in/check-out, mobile-compatible interfaces for yard teams, automated alerts for overdue maintenance or certification expiry.

Redesigned the Ariba procurement flow to enforce approval hierarchies and capture vendor performance data. Rebuilt CapEx planning templates against live AM depreciation schedules. Trained 200+ users across technical, logistics, and finance functions.

The Outcome

Full traceability achieved for 5,000+ assets across six countries — location, maintenance status, certification, and depreciation all visible in one system. Cross-border movement compliance improved significantly, reducing regulatory risk during customs inspections.

CapEx planning cycles shortened from 3 months to 3 weeks because finance could now pull live asset data instead of chasing spreadsheets from each department. Procurement cycle times reduced through Ariba automation, with vendor spend becoming auditable for the first time.

Stack SAP AM SAP MM SAP PM SAP Ariba SAP Logistics C#/.NET RFID/Barcode SQL Server AutoCAD Integration
SEC
Penetration Testing · ICC Compliance · VIP Data
CASE STUDY / 04

Cybersecurity Audit —
BLS Events / First Arab Bank

High-profile event management company handling VIP data for financial institution clients. First Arab Bank required ICC compliance before the client relationship could continue. The audit had to find and close every gap — reputational stakes were absolute.

Client
BLS Events
End-Client
First Arab Bank
Standard
ICC Compliance
Result
Full Compliance
ICC
Compliance Achieved
0
Critical Findings Unresolved
VIP
Data Protection Verified
The Problem

BLS Events managed guest lists, personal details, and sensitive information for high-profile banking events hosted by First Arab Bank. When First Arab Bank's security team reviewed the arrangement, they flagged gaps in how VIP data was being stored, transmitted, and accessed by event management staff.

ICC compliance was the threshold — without it, the contract would not continue. The deadline was fixed, the stakes were real, and a surface-level audit report would not pass First Arab Bank's internal review.

The Approach

Conducted a full penetration test and vulnerability assessment of BLS Events' systems — covering web applications, internal network, data storage practices, and access control systems. Went beyond automated scanning to manually verify privilege escalation paths, encryption in transit and at rest, API authentication, and physical access controls at event venues.

Mapped every point where VIP personal data touched external systems or third-party integrations. Produced a remediation roadmap with severity-tiered findings, owner assignments, and verification checkpoints.

The Outcome

All critical and high-severity findings remediated within the compliance timeline. ICC certification achieved. First Arab Bank's security team signed off on the audit report and the client relationship continued.

Beyond compliance, the engagement surfaced architectural improvements: unified access logging, encrypted event data export workflows, and a formal data retention and destruction schedule that hadn't previously existed.

Stack Penetration Testing OWASP Top 10 Network Analysis API Security Access Control Review Encryption Audit ICC Compliance ISO 9001
VSOP
VSOP87 · SIMD · Relativistic Corrections · Rust
CASE STUDY / 05

Planetary Ephemeris Engine
— Sub-Arcsecond Precision

Astronomical position calculation at the accuracy demanded by observatories and precision instrumentation — not the approximations that consumer apps use. VSOP87 series implemented in Rust with SIMD acceleration and full relativistic corrections.

Domain
Scientific Computing
Algorithm
VSOP87
Accuracy
Sub-arcsecond
Language
Rust + SIMD
<1"
Positional Accuracy (arcsec)
8
Planets Computed Simultaneously
SIMD
Hardware Vectorization
The Problem

Most astronomical position libraries available as open source are either Python-based (too slow for real-time instrumentation) or use simplified models that introduce errors of several arcminutes over long time spans. Applications requiring precision — telescope mount controllers, satellite tracking, religious calendar calculations, navigation systems — cannot tolerate these approximations.

The challenge: implement the full VSOP87 planetary theory with correct aberration, nutation, precession, and relativistic light-time corrections, while achieving performance suitable for real-time or high-throughput computation.

The Approach

Implemented the full VSOP87 series (heliocentric elliptic) in Rust — covering all eight planets from Mercury to Neptune, with separate L, B, R series for ecliptic longitude, latitude, and radius vector. Applied IAU 2006 precession theory and IAU 2000B nutation model, plus aberration correction using the full Newcomb series.

Vectorized the inner summation loops using portable SIMD intrinsics, allowing multiple terms to be evaluated simultaneously. Relativistic light-time iteration converges to single-precision limit in under 4 passes for all solar system bodies. Full ephemeris for all planets computes in under 100µs on modern hardware.

The Outcome

Sub-arcsecond accuracy verified against JPL Horizons reference positions for dates spanning 2000 BCE to 3000 CE. The engine handles coordinate frame transforms (ecliptic → equatorial → horizontal), topocentric corrections for any observer location, and Julian/Gregorian calendar conversions.

The SIMD implementation delivers 8–12× throughput over scalar implementation on the same hardware, enabling use cases that require computing thousands of positions per second — trajectory planning, occultation prediction, real-time mount control.

Stack Rust SIMD (portable_simd) VSOP87 IAU 2006 Precession IAU 2000B Nutation Relativistic Corrections JPL Horizons Verification Coordinate Transforms
AISHA
Self-Healing Automation · CV Recognition · Rust/WASM
CASE STUDY / 06

AISHA — Intelligent
Web Automation

Conventional automation breaks the moment a website changes its layout. AISHA is different: it understands what it's looking at through computer vision, heals itself when selectors break, and mimics human interaction patterns to avoid detection.

Domain
Web Automation
Approach
CV + Self-Healing
Runtime
Rust / WASM
Fingerprint
Human-Pattern Evasion
CV
Vision-Based Element Recognition
Self
Healing on Layout Changes
0
Bot Detection Triggers (Tested)
The Problem

Enterprise data gathering from web sources fails constantly. CSS selectors break on redesigns. XPath fragments on DOM structure that changes with A/B tests. IP-based rate limiting blocks naive crawlers. Bot detection systems flag anything that doesn't look human. Organizations need automation that can survive in the real web environment — unreliable, adversarial, constantly changing.

Existing tools like Selenium and Playwright work for controlled environments but require constant maintenance when deployed against production sites. The selector maintenance burden alone can exceed the value delivered.

The Approach

Built AISHA with a fundamentally different architecture: computer vision as the primary locator, not DOM selectors. Elements are identified by visual appearance and spatial relationship — the way a human would locate a button — not by brittle attribute chains. When a visual pattern fails, the system uses semantic understanding of page context to find the nearest functional equivalent.

Interaction timing uses statistical models of human mouse movement and keystroke cadence. Fingerprint avoidance goes beyond user agent spoofing: canvas fingerprint randomization, WebGL parameter masking, font enumeration control, timing entropy injection — all implemented in Rust compiled to WASM.

The Outcome

AISHA maintained successful operation through multiple complete redesigns of target web applications without code changes — the self-healing CV layer adapted automatically. Zero bot detection triggers across all tested deployments using standard commercial bot detection services as the benchmark.

The WASM compilation target enables browser-side execution where needed, or server-side orchestration for high-throughput scenarios. The architecture supports use cases from enterprise data collection to automated testing of applications that are resistant to conventional tools.

Stack Rust WebAssembly Computer Vision Self-Healing Selectors Fingerprint Avoidance Human-Pattern Simulation Canvas/WebGL Masking Chromium DevTools Protocol
BRING US THE PROBLEM

Every project here started with
someone saying "we don't think this is possible."

Discuss Your Problem →