Production AI Systems

All solo-built, deployed, and serving real users — each solving a distinct set of enterprise problems.

SF-Assistant

Conversational AI Agent for SAP SuccessFactors

Production

A 28-tool, 35-iteration agentic AI system that lets HR teams query employee data, troubleshoot permissions, diagnose workflow issues, and analyze org structures via natural language — replacing hours of manual SAP investigation with conversational answers.

Solo Builder — serving ~50 client companies in production

Python 3.13FastAPIAWS BedrockClaude Sonnet 4.5SeleniumSQLiteFAISSBM25Streamlit

28

Tools

30+

OData Entities

2

Microservices

Key Features

  • 28 AI-accessible tools across 5 domain modules (Employee, RBP/Roles, Query, Permissions, Automation)
  • 35-iteration agentic loop with custom caching, retry logic, and response validation
  • ~600-line system prompt with 16 embedded few-shot examples
  • Multi-browser session management with per-tenant isolation
  • +4 more

Technical Challenges Solved

  • Two-service split specifically to isolate slow Selenium sessions (10–40s) from the AI chat service, enable independent scaling, and contain crashes
  • Per-tenant SAP session caching with thread-safe double-checked locking and auto-eviction/re-login

SFCodeBot

Multi-Service Enterprise AI Analytics Platform

Production

A separate multi-service platform that turns raw test-automation logs and arbitrary Excel workbooks into natural-language-queryable, AI-analyzed data — with a hybrid FAISS+BM25+Claude retrieval stack and automatic model failover.

Solo Builder — serving ~50 client companies in production

PythonFastAPIFlaskStreamlitDuckDBPandasPolarsPlotlyFAISSBM25AWS BedrockCohere Embed v4

4

AI Subsystems

15+

Chart Types

8

Pipeline Stages

Key Features

  • 4 AI subsystems: Log Analysis, Excel Analytics, Data Manipulation, Knowledge-Base Generation
  • 8-stage log-analysis pipeline: fetch → segment → preprocess → context → dedup → chunk → AI analyze → summary
  • Hybrid retrieval: FAISS + BM25 + Claude re-ranking for finding the right sheet among hundreds
  • 4-stage Model Knowledge Generator that auto-infers foreign-key relationships via value-overlap
  • +4 more

Technical Challenges Solved

  • Safe execution of LLM-generated code at production scale with automated self-repair
  • Multi-model failover with full cost tracking across all call types
Visit live site

ExcelSense-AI

Natural Language Excel Analytics Platform

Production

A Streamlit-based analytics platform enabling natural language queries across 50+ Excel sheets with real-time AWS S3 and Google Sheets synchronization. Processes datasets up to 100MB with sub-3-second response times.

Solo Builder

PythonStreamlitPandasAWS S3Google Sheets APIDockerAWS EC2

<3s

Response Time

50+

Sheet Capacity

60%

Cost Reduction

Key Features

  • Natural language to optimized Pandas operations with comprehensive audit trails
  • Real-time sync with AWS S3 and Google Sheets
  • Processes datasets up to 100MB with <3s response times
  • Intelligent caching reducing per-query costs by 60%
  • +2 more

Technical Challenges Solved

  • Query translation from natural language to optimized Pandas operations
  • Real-time synchronization across S3 and Google Sheets

PakLaw AI V3

Bilingual Legal AI Platform

Production

A legal-tech platform with lawyer directory, RAG-based semantic search across 10,000+ Pakistani legal documents, in-app communication, and full English/Urdu bilingual support. Optimized to run on under 4GB RAM while achieving 92% OCR accuracy on complex multi-column court judgments.

Solo Builder — live with real user traffic

FastAPIReactTypeScriptSQLiteChromaDBRedisSentence TransformersTesseract OCR

10,000+

Documents

<500ms

Query Latency

92%

OCR Accuracy

Key Features

  • Semantic search across 10,000+ Pakistani legal documents with <500ms average latency
  • Custom Urdu OCR preprocessing pipeline achieving 92% accuracy on complex court judgments
  • Hybrid BM25 + dense embedding retrieval improving precedent matching by 35%
  • Memory-optimized: reduced RAM from 16GB to 4GB through model quantization and efficient indexing
  • +2 more

Technical Challenges Solved

  • Reducing RAM from 16GB to under 4GB through model quantization while maintaining accuracy
  • Custom Urdu preprocessing for Tesseract OCR on complex multi-column layouts
Visit live site

Ready to Talk?

Get clarity on where AI can actually improve your business.

Start with a free 30-minute strategy call.