Case Study 01
AI & Me: A Year of Growth
Client: Andovar, Lxt AI, Lablebox, Outlier, Uber
Role: AI Training Specialist · Data Annotator · Prompt Engineer · Multilingual QA
Duration: 18+ Months
Client: Andovar, Lxt AI, Lablebox, Outlier, Uber
Role: AI Training Specialist · Data Annotator · Prompt Engineer · Multilingual QA
Duration: 18+ Months
This case study documents my hands-on journey working inside real-world AI training pipelines across global platforms. Over 18+ months, I contributed to building, refining, and validating datasets that power modern NLP, computer vision, speech recognition, and multimodal AI systems.
Rather than working on a single product, this experience spans multiple AI production environments, offering a rare, ground-level view of how data quality, annotation rigor, and human judgment directly shape model behavior.
Modern AI systems fail not because of algorithms alone but because of noisy data, weak labels, cultural blind spots, and poorly designed prompts.
Across industries (autonomous systems, voice assistants, chatbots, OCR, recommendations), teams face recurring challenges:
Inconsistent annotations across large datasets
Bias and hallucinations in language models
Poor multilingual and accent handling
Low-quality speech-to-text and OCR outputs
Weak prompt logic leading to unreliable reasoning
My role was to operate inside these fault lines identify weaknesses early and improve data fidelity before models reached production scale.
I worked as a distributed AI contributor, collaborating with operations leads, QA reviewers, and automated validation systems. My responsibilities evolved from execution-heavy tasks to higher-order evaluation and refinement.
Core responsibilities included:
Image, video, audio, and text annotation
Prompt engineering and response evaluation
Multilingual localization and QA (English, Hindi, Punjabi)
Speech-to-text correction and voice dataset collection
Bias detection, hallucination spotting, and output validation
Objective: Improve speech recognition and NLP performance across diverse accents and languages.
What I did:
Recorded high-quality multilingual voice commands
Edited and corrected AI-generated transcripts
Labeled speaker turns, corrected misrecognitions, and cleaned noisy audio outputs
Impact:
Reduced transcription errors in real-world audio
Improved model robustness for accented and conversational speech
Objective: Train vision models for autonomous systems, object detection, and human motion analysis.
What I did:
Annotated large-scale image and video datasets using bounding boxes, segmentation, and tags
Labeled vehicles, objects, traffic elements, and human joints
Designed skeleton maps for pose estimation and motion tracking
Impact:
Strengthened model accuracy for real-time decision-making
Enabled downstream applications in robotics, sports analytics, and safety systems
Objective: Improve reasoning quality, multilingual fluency, and logical consistency in LLM outputs.
What I did:
Designed prompts in English, Hindi, and Punjabi
Built mathematical and logical reasoning prompts
Rated model responses for accuracy, tone, intent, and reasoning depth
Flagged hallucinations, bias, and incomplete reasoning
Impact:
Improved step-by-step reasoning quality
Reduced hallucinations and culturally incorrect outputs
Objective: Train multilingual NLP systems with culturally accurate data.
What I did:
Translated English content into Punjabi with cultural and contextual precision
Validated localization quality for AI training pipelines
Completed structured AI annotation and QA workflows
Impact:
Improved regional language understanding in virtual assistants
Reduced semantic drift in translated datasets
Objective: Improve chatbot reliability and sentiment accuracy.
What I did:
Wrote and refined prompts in English and Hindi
Labeled outputs for sentiment, relevance, and correctness
Performed fine-grained QA to reduce bias and inconsistency
Impact:
Increased response coherence and human-likeness
Strengthened sentiment classification reliability
Objective: Power recommendation systems, vision AI, and document automation.
What I did:
Classified text into domains (Tech, Business, Politics, Entertainment, Sports)
Labeled food images for vision-based restaurant and dietary apps
Performed object counting and handwriting recognition for OCR
Rated Hindi prompts and responses for linguistic quality
Impact:
Improved recommendation accuracy
Enhanced OCR reliability for handwritten data
Strengthened multilingual response quality
Annotation Platforms: Darwin V7 Labs, LDP, proprietary QA tools
Modalities: Image, Video, Audio, Text
Languages: English, Hindi, Punjabi
Skills: Prompt engineering, speech QA, OCR validation, localization, bias detection
What this experience taught me:
AI quality is a data problem before it is a model problem
Small annotation decisions compound at scale
Multilingual and cultural context is non-negotiable for global AI
Human judgment remains critical in closing the gap between raw models and usable systems
This wasn’t theoretical AI work.
It was production-grade, high-stakes, detail-heavy work the kind that determines whether AI systems behave responsibly, accurately, and inclusively.
I bring:
Execution depth
Systems thinking
Multilingual intelligence
A strong ethical lens toward AI development
Client: Unlocking Venture Capital Podcast (VenturX Capital, New York)
Role: Marketing Intern
Duration: 4 months
This case study captures my work scaling a founder-led media brand through design-driven, data-informed marketing -without paid amplification.
Unlocking Venture Capital is a niche media platform spotlighting underrepresented and emerging voices in the venture capital ecosystem. The brand’s strength lay in thoughtful, long-form conversations. The challenge was translating that depth into consistent, high-retention, multi-platform content, while preserving the founder’s authentic voice.
When I joined, the problem wasn’t content volume, it was structure and signal.
Key challenges included:
Inconsistent visual identity across YouTube and LinkedIn
Strong ideas, but weak hooks and retention in early content seconds
Limited clarity on what was driving engagement vs. vanity metrics
Heavy founder dependency with no repeatable content system
The goal was to bring clarity, consistency, and compounding growth to a lean, founder-driven operation.
Improve content consistency and visual identity across platforms
Increase organic reach and engagement (no paid promotion)
Build a repeatable, low-friction content production system
Introduce data-driven reporting to guide weekly iteration
I worked directly with the Founder across creative production, experimentation, and performance analysis.
My responsibilities spanned:
Podcast video editing and high-retention short-form cuts
Thumbnail, post design, and colour-based branding experiments
Content ideation, scripting, and campaign analysis
Multi-channel performance tracking and insight reporting
Designing repeatable workflows for sustainable production
This role sat at the intersection of design, content, and analytics.
Before scaling output, I focused on improving how content appeared in-feed.
Actions taken:
Designed experimental thumbnails to test click and retention behavior
Introduced colour-led branding systems for visual consistency
Created post creatives aligned with episode themes and guest profiles
Outcome: Improved visual clarity and brand recognition across crowded feeds—making the content easier to identify and trust before the click.
Rather than optimizing for impressions alone, the strategy centered on watch quality.
Actions taken:
Edited long-form episodes into tighter, high-retention formats
Tested hooks, pacing, and visual emphasis in the first seconds
Used watch time per view as the primary optimization signal
Results:
2,062 total views
30 hours of watch time
Average view duration: 2:47
This indicated strong content quality once viewers clicked.
LinkedIn was positioned as the primary professional discovery channel.
Actions taken:
Planned and published consistent, insight-led posts
Balanced founder voice with clean, design-forward presentation
Monitored organic reach and engagement trends weekly
Results (100% organic):
5,073 total impressions
783.8% growth in impressions over 4 months
Engagement: 122 reactions · 31 comments · 14 reposts
To support iteration, I introduced lightweight but actionable reporting.
Actions taken:
Consolidated YouTube and LinkedIn data into readable summaries
Tracked retention, reach, and engagement patterns
Shifted review conversations from “what performed” to “why it performed”
This enabled faster, clearer weekly decisions.
To reduce founder dependency and ad-hoc execution:
Designed workflows from ideation → production → distribution
Documented repeatable formats for thumbnails, posts, and reporting
Outlined automation scenarios for future scaling
The result was a portable content engine, not just one-off execution.
7,000+ total impressions across platforms
45+ hours of audience engagement
Improved content consistency and viewer experience
Stronger top-of-funnel reach with clear retention signals
A documented, scalable system for ongoing content production
(Hosted on Notion - intentionally curated, not overloaded)
Podcast Production & Analysis (Episodes 01–04)
Audience & Content Performance Analysis
4-Month Marketing Growth Report
Campaign Analysis & Datasets
Podcast Ideation Vault
Automation Scenarios for Content Operations
Design clarity compounds distribution
Retention is a stronger signal than reach
Founder-led brands scale best with systems not volume
Data is most useful when it explains behavior, not just performance
This project demonstrates my ability to operate across design, content, and analytics, building growth systems that scale without losing voice or intent.
Client: Figmenta (UK)
Role: Content & QA (Product, Operations & AI Intern (Expanded Scope)
Duration: 3 Months
This case study documents my experience operating inside a founder-led digital agency during a period of structural transition.
Figmenta was actively shifting toward a project-based, AI-first operating model, while simultaneously managing legacy initiatives, evolving priorities, and live client work. I joined during this transition, initially in a content and operations capacity but the role quickly expanded into product analysis, AI systems, frontend collaboration, and project ownership.
Rather than solving a single defined problem, the work required navigating ambiguity, speed, and incomplete information, while helping leadership bring clarity, momentum, and sustainability to multiple parallel initiatives.
There was no single problem statement.
Instead, the environment presented a set of real operational realities:
Projects at mixed stages of maturity, including initiatives paused for 1–2+ years
Ambiguous ownership and fluid role boundaries
High expectations for independent judgment and speed
Founder-led decision-making with limited process guardrails
An active transition from traditional team structures to a Project-Based Organization (PBO)
The core challenge was not execution alone but operating effectively under ambiguity while maintaining quality, alignment, and forward momentum.
Over the course of two months, my role evolved from execution support to problem ownership and decision support.
Depending on project needs, I operated across multiple capacities:
Project Owner on internal and early-stage product initiatives
Product / Growth Analyst supporting leadership decisions
AI workflow builder for content, SEO/AEO, and operations
Cross-functional collaborator across engineering, design, content, and leadership
Interview panel participant for new hires
Rather than being confined to a single function, I was trusted to frame problems, surface risks, and propose solutions across domains.
Took ownership of active and dormant initiatives, including projects paused for over two years
Conducted product feasibility, competitor analysis, and technical sustainability assessments
Flagged design and architectural risks early to support go / no-go decisions
Provided structured analysis under time and information constraints
Designed and implemented production-oriented AI workflows, including:
Content generation and visual ideation systems
SEO and AEO research pipelines
Operational automation for repetitive internal tasks
AI was treated not as experimentation, but as a reliable support layer for scale and consistency.
Conducted deep keyword and search intent analysis using structured datasets
Built AEO-focused playbooks aligned with AI-driven search behavior
Translated insights into clear, actionable recommendations adopted by the team
Transitioned from backend analysis into frontend and visual collaboration
Worked directly in Figma with designers on layout, typography, and direction
Produced visual concepts that influenced live design decisions
Participated in candidate interviews alongside Operations and Efficiency Leads
Evaluated candidates on both skill and mindset
Contributed to hiring decisions in a growing team
Across projects, my approach remained consistent:
Clarify the real problem (not just the task)
Identify constraints - technical, organizational, or time-based
Operate independently, but communicate decisively when boundaries mattered
Avoid overproduction; prioritize clarity, sustainability, and impact
Treat failure and resets as system signals, not personal setbacks
This helped stabilize projects under pressure and supported clearer leadership decisions.
Multiple initiatives progressed from ambiguity to structured direction
Leadership adopted several product, SEO, and AEO recommendations
AI workflows reduced manual effort and improved operational consistency
Design and product decisions were informed earlier by feasibility analysis
Trust expanded from execution tasks to ownership and evaluation responsibilities
Most importantly, my role shifted from "doing assigned work" to being consulted for perspective, signaling deeper integration into decision-making processes.
Ambiguity is manageable when framed correctly
Ownership is earned through judgment, not titles
Speed matters—but sustainability matters more
Silence creates friction; clarity prevents it
Strong systems outperform heroic effort
This experience clarified the environments where I perform best: fast-moving, founder-led teams that value autonomy, analytical depth, and cross-functional thinking.
This was not a linear internship it was an immersion into how real organizations think, decide, and adapt under pressure.
The work pushed me beyond execution into problem ownership, sharpening my product sense, communication, and ability to operate under uncertainty. That shift will continue to shape how I approach future roles in product, strategy, and AI-driven operations.
Client: Figmenta.com
Role: SEO Operations & Systems Owner
Phase: Month 3 of Internship
Focus: Technical SEO, Indexing Hygiene, Content Ops
This case study documents a specialized SEO operations intervention focused on restoring indexing health, crawl efficiency, and long-term search stability for Figmenta.com.
When I stepped into SEO operations, the site was experiencing severe indexing instability caused by legacy URL debt, stale pages, and an external security incident that polluted Google Search Console with unauthorized sitemaps.
The goal was not short-term rankings but structural cleanup, signal normalization, and sustainable crawl behavior.
At the start of my SEO operations work:
~4,000+ stale URLs existed in the background (legacy, invalid, or abandoned)
Total live pages: ~504
Indexed pages: ~250 (~50%)
Major incident: ~1,000 unauthorized XML sitemaps submitted by an external source
Indexing signals were noisy, inconsistent, and unreliable
The system problem was clear: Google could not reliably understand what mattered on the site.
Polluted sitemap environment damaging trust signals
Large volume of invalid and stale URLs diluting crawl budget
Inconsistent indexing feedback due to mixed signals
Risk of stale URLs continuously recirculating
No clear operational framework for ongoing SEO hygiene
This was not a content problem alone; t was a crawl and indexing systems problem.
I treated this as an SEO operations cleanup, not a one-off technical fix.
First priority: eliminate active damage.
Removed all ~1,000 unauthorized sitemaps from Google Search Console
Stabilized the sitemap environment to prevent further signal pollution
This step alone reduced noise and stopped compounding indexing errors.
Once the environment was stable:
Performed structured URL inspection on priority pages
Actively requested indexing for valid, high-value URLs
Monitored indexing responses to validate crawl behavior
The focus was on precision, not bulk actions.
Over time:
Invalid and unnecessary URLs were systematically reduced
Indexing signals began to normalize
Crawl consistency improved across the site
This was measured not by rankings but by index coverage quality and feedback clarity.
As of the latest report:
Indexed pages: ~560 (nearly full coverage of all valid pages)
Not indexed pages: 368 remaining
Legacy stale URLs: still present but actively declining through validation
Indexing errors significantly reduced
Crawl behavior stabilized and predictable
The site has effectively moved from indexing instability → controlled normalization.
Invalid and stale URLs reduced significantly
Indexing signals cleaned and normalized
Crawl efficiency improved
Crawl relevance strengthened
Faster and clearer diagnosis of indexing issues
This created a strong technical foundation for future rankings, impressions, and content quality gains.
Cleaner indexing directly enables:
Better crawl budget allocation
Higher relevance signals per page
Faster search engine understanding of site structure
Stronger long-term visibility and ranking potential
Reduced operational SEO debt
This work ensures that future content efforts compound instead of being diluted.
Alongside technical SEO cleanup, I also supported content execution:
Wrote and optimized SEO-focused articles
Created and edited article visuals using Canva and Adobe tools
Built a custom GPT-based image generation board to produce consistent, abstract visual assets aligned with brand tone
This system ensured visual consistency at scale while reducing manual design effort.
Gradual removal of irrelevant and legacy URLs
Prevent stale pages from re-entering the index
Maintain long-term indexing cleanliness
Continue aligning content creation with crawl relevance
This transitions SEO from reactive cleanup to stable, operational maintenance.
Indexing health is a systems problem, not a one-time fix
Crawl clarity enables faster growth later
SEO operations benefit from calm, methodical execution
Cleaning foundations first multiplies future content ROI
This case study demonstrates my ability to manage SEO as an operational system, balancing technical cleanup, content execution, and long-term search stability under real-world constraints.
Platform: Personal LinkedIn Account
Tool: Waalaxy
Duration: 15 Days (Campaign <50% complete)
This case study documents a short, controlled experiment to test whether automation-driven outreach can outperform manual activity on LinkedIn in terms of growth, efficiency, and conversation quality.
The goal was not vanity growth, but to validate:
Can structured automation accelerate reach?
Can it maintain meaningful engagement?
How much time can be saved vs manual effort?
Followers: ~700
Profile Views (recent): ~30
Growth driven primarily through manual activity and organic posting
Used a Waalaxy automation sequence:
LinkedIn connection requests (targeted)
Follow-up email outreach
Approach focused on:
Consistent daily execution (via automation)
Clean, non-spammy messaging
Allowing conversations to evolve naturally after connection
Followers: 700 → 905
→ +205 growth (~29% increase)
Profile Views: 30 → 154
→ +124 increase (~413% growth)
Time Saved: ~10–15 hours over 15 days
→ Equivalent to ~40–60 minutes/day of manual effort avoided
Campaign is still under 50% completion, with growth continuing
Automation created consistent daily reach without effort spikes
Higher profile visibility led to organic discovery beyond outreach
Conversations initiated through campaigns were meaningful, not shallow
Reduced reliance on constant platform activity for growth
Small, consistent automation > irregular manual effort
Visibility compounds quickly when outreach + profile clarity align
Automation is most effective when paired with clean positioning and content
Growth doesn’t require constant presence—just structured systems
Short time window (15 days)
Early-stage results (campaign not fully completed)
Growth quality depends heavily on targeting and messaging
Complete full campaign cycle
Refine targeting segments
Test variations in messaging
Measure long-term engagement vs short-term growth
This experiment shows that automation, when used thoughtfully, can significantly accelerate LinkedIn growth while saving time-without sacrificing conversation quality.
The real leverage came from consistency, not intensity.