SharabhaTech — Powered by Isana Systems
Case Studies

Real AI systems. Real measurable impact.

Every project below ran in production for a real client, on real data, with quantified business outcomes. This is what "outcome-first AI" looks like in practice.

Featured Projects

Ten production AI systems with measurable outcomes

Selected detail from our delivery track record. Each case study is a real client engagement run end-to-end on real data.

Project 01 · Developer Productivity

AI Code Review Agent for Engineering Productivity

Challenge

Senior engineers and architects were spending ~50% of their time on code reviews, yet pull and merge requests still waited hours to days for the first review. Architectural concerns and performance & scalability issues regularly slipped through manual reviews under that load.

AI Solution

  • AI agent posts the first review comment on every PR/MR within minutes
  • Detects functional bugs, architectural issues, and performance bottlenecks
  • AI-as-judge layer suppresses nit-picks and review noise — only signal surfaces to humans
  • Human feedback on agent comments tunes the model to repo-specific conventions
First review comment in ~4 minutes average · Senior engineer review time reduced 60% · PR/MR merge cycle time reduced 65%
Tech: Claude API · Claude Code · Python · Git / GitHub / GitLab APIs · CI/CD Integration
Project 02 · AdTech / Revenue Operations

AdTech Revenue Insight, Alerting & Correction Agent

Challenge

Daily ad revenue had to be reconciled across multiple time granularities (DoD, WoW, MoM, QoQ, plus WTD / MTD / QTD) and dimensions (device, revenue property, geography). Manual analysis was slow, business teams were blocked behind data engineers, and revenue leakage went uncaught for hours.

AI Solution

  • Data engine ingests daily revenue JSON into a PostgreSQL pgvector RAG store
  • Revenue Analysis Engine replaces manual reconciliation across every time-cut and dimension
  • Alerts on revenue drops and gains within hours of day-boundary
  • Auto-triggers corrective campaign actions — re-balancing device/property priorities to recover revenue
  • Natural-language chatbot lets business users query revenue trends without data-engineering tickets
Zero manual effort for revenue-loss analysis · Reports & alerts within 1–2 hours of day boundary · Corrective actions queued before office hours · Daily email digest + on-demand chatbot
Tech: Claude API · PostgreSQL pgvector · RAG · Python · JSON ingestion · Email & Chatbot delivery
Project 03 · FMCG / Sales Operations

FMCG Sales Insight, Alerting & Correction Agent

Challenge

Daily sales data spanning sales managers, agencies, products, and product categories was reviewed manually. Sales drops took days to surface, pricing and promotion decisions ran on stale data, and business users had no ad-hoc path to answers without data-engineering support.

AI Solution

  • Data engine ingests daily sales JSON into a PostgreSQL pgvector RAG store
  • Sales Analysis Engine replaces manual reconciliation across every dimension
  • Alerts on sales drops and gains within hours of day-boundary
  • Auto-triggers corrective actions — add/withdraw products or categories, apply daily discounts or combo offers
  • Natural-language chatbot for business users to query sales trends in plain English
Zero manual effort for sales-loss analysis · Reports & alerts within 1–2 hours of day boundary · Corrective actions queued before office hours · Daily email digest + on-demand chatbot
Tech: Claude API · PostgreSQL pgvector · RAG · Python · JSON ingestion · Email & Chatbot delivery
Project 04 · IT Support Operations

Zendesk Support Ticket Root Cause Analysis

Challenge

2,000+ support tickets with no systematic visibility into root causes. Manual categorisation was inconsistent, recurring incidents went undetected, and engineering had no data-driven prioritisation framework.

AI Solution

  • Claude AI analysed every ticket via Zendesk REST API
  • Auto-classified root causes into structured categories
  • Pattern detection engine identified systemic issues
  • Generated 27 user stories mapped to defects with effort estimates
25% projected support ticket reduction · 2,000+ tickets → 27 actionable user stories · 14-week implementation roadmap delivered
Tech: Python · Claude API · Zendesk REST API · Excel · Pattern Analysis
Project 05 · HR Tech / Talent Acquisition

AI-Powered Video Assessment & Proctoring

Challenge

Manual review of 15–30 minute developer demo videos created a hiring bottleneck with subjective scoring and no integrity monitoring for remote assessments.

AI Solution

  • Claude Vision API analyses video frames + audio transcripts
  • 5-dimension scoring: Communication, Technical Depth, Code Quality, Functionality, Problem-Solving
  • Proctoring engine flags mobile use, eye patterns, scripted answers
  • Timestamped integrity incident reports
90%+ reduction in manual review time · Consistent bias-free scoring · ~$2–7 per video · Scalable to any candidate volume
Tech: Claude Vision · OpenCV · FFmpeg · Anthropic SDK · Python
Project 06 · HR Tech / Recruitment

AI-Assisted Candidate Evaluation Framework

Challenge

No standardised rubric for evaluating .NET developers. Resume reviews took 45–60 minutes with inconsistent standards and no objective fitment calculation.

AI Solution

  • Weighted skills framework: 6 categories, 0–5 scale
  • Claude maps CV content to JD requirements automatically
  • Mandatory filter gates (degree, core language experience)
  • Auto-calculated fitment % with tiered ratings
Resume review 60 min → under 10 min · Objective fitment scoring · Reusable across unlimited future roles
Tech: Python · Claude API · Excel · Playwright · Markdown
Project 07 · Customer Experience

Automated Customer Sentiment Analysis

Challenge

No visibility into customer sentiment across 12 months of interactions. Churn-risk customers were identified only after escalation, and manual ticket review was impractical at scale.

AI Solution

  • Claude AI analysed all Zendesk tickets across 12 months
  • Multi-dimensional output: sentiment, emotion, urgency, churn-risk
  • Sentiment score –10 to +10 with emotion taxonomy
  • Monthly CSV/JSON files with annual consolidation
Full-year sentiment dataset delivered automatically · Churn-risk identified before escalation · $0.50–$2 per 100 tickets analysed
Tech: Python · Claude API · Zendesk REST API · Excel · CSV/JSON
Project 08 · AI Operations / DevOps

AI Engineering Standardisation & SOP

Challenge

Claude Code sessions were unstructured and unmeasurable. No way to trace logs back to tasks, no benchmarking of AI productivity across engineers, no replicable workflows.

AI Solution

  • Task ID system (PROJECT-NNN) for every AI session
  • Structured session protocols: ##TASK / ##RESUME / ##SCOPE-ADD
  • Prompt templates with mandatory context fields
  • Centralised Task Registry with cross-team metrics
Full audit trail per AI session · Engineer prompt-quality benchmarking · Reusable prompt library accelerates future projects
Tech: Claude Code · Python · JSONL · Markdown · Git
Project 09 · Government ERP / Software Maintenance

AI-Powered Legacy ERP Bug Resolution Pipeline

Challenge

Legacy VB.NET & PHP ERP serving US govt agencies — no documentation, retired original developers, India team without domain knowledge. Bug fixes took 1–2 weeks with frequent regressions.

AI Solution

  • Claude Code reads the escalated Zendesk ticket via API
  • Identifies the affected module & analyses source code
  • Posts root cause + diff + unit tests + regression suite to Jira
  • Developer reviews, approves; Jenkins handles deployment
60–75% faster bug fix cycles (1–2 weeks → 2–4 days) · Significant drop in production regressions · Developers shift from writing to reviewing
Tech: Claude Code · Zendesk API · Jira API · Jenkins · VB.NET · PHP
Project 10 · Knowledge Management / Support Ops

AI-Powered Support Knowledge Base Builder

Challenge

~100 ERP support tickets per week with all resolution knowledge trapped in TSE phone calls. Zero institutional knowledge base — every issue investigated from scratch, no audit trail for RCA.

AI Solution

  • Helpline moved to MS Teams with call recording & transcripts
  • TSE states Zendesk ticket # aloud at start of each call
  • Daily automation pulls transcripts, links each to its ticket
  • Claude API distils into structured KB entry on the ticket
100% of ~100 weekly support interactions captured as structured knowledge · TSE onboarding time slashed · Directly unlocked downstream RCA
Tech: Claude Code · MS Teams API · Zendesk API · Python · Claude API

Have a problem like one of these?

If any of the challenges above sound familiar, let's talk. We can usually tell within a 30-minute call whether AI is the right fit and roughly what it would take.

Book a Discovery Call