Real problems. Applied AI. Engineering that matters.

The challenge

Parents want calm, age-appropriate bedtime stories without searching or configuring anything. Fixed story libraries run out quickly and never feature the child's own world.

What we built

Datu generates a new illustrated, narrated bedtime story for each child, every night. A recurring cast of characters, Datu the baby elephant, Rio the explorer, Orion the young scientist, Dr Maya and Coach Ray, carries gentle lessons in kindness, courage, empathy and teamwork.

How it works

1. Personalized narratives adapted to the child's age, interests and the values parents want to reinforce.
2. Character and story continuity so familiar characters and themes carry across nights.
3. AI-generated illustrations in a consistent visual style.
4. Natural narration through text-to-speech, ready for lights-out.
5. Multilingual stories for families who read in more than one language.
6. Safety and quality checks on every story for age-appropriateness and tone.

Engineering considerations

Coordinating text, image and voice models within a few seconds; keeping per-story inference cost low enough for a consumer price; consistent character art; content safety for young audiences.

What it demonstrates

We can take several AI models beyond demos and ship them as one product that people use every day.

The challenge

In video commerce, shoppers discover products through a feed instead of a search box. The feed has to feel relevant, fresh and varied while every video stays connected to something people can buy.

Our role

Our team led product and launch for video-commerce initiatives at Flipkart, one of India's largest marketplaces, spanning product development, content operations, marketplace execution and recommendation-driven discovery.

How it works

1. Behavior signals: views, watch time, product clicks and session context.
2. Content and product representation: metadata and learned features linking each video to its products.
3. Candidate retrieval from behavior, item relationships and content similarity.
4. Personalized ranking on predicted relevance and engagement, within availability and merchandising rules.
5. Feed composition that balances relevance with variety, freshness and discovery.
6. Experimentation through controlled tests on discovery and shopping metrics.

Engineering considerations

Cold start for new videos and users, a fast-changing catalog, real-time session signals and low-latency serving.

The challenge

Markets react to narratives that spread on social media. Keyword tracking cannot tell meaningful sentiment from noise, and financial posts are full of ambiguous tickers, sarcasm and posts covering several assets at once.

What we built

An NLP pipeline that turns Twitter/X conversations into entity-specific market intelligence.

How it works

1. Ingestion and cleaning of permitted social data, with deduplication and spam filtering.
2. Financial entity recognition that resolves companies, tickers and instruments, including ambiguous symbols.
3. Entity-level sentiment so one post can be bullish on one stock and bearish on another.
4. Narrative detection that clusters related conversations into emerging themes.
5. Time-series aggregation of sentiment and conversation volume per entity.
6. Analyst interface with search, charts and alerts.

Engineering considerations

Domain vocabulary, sarcasm, coordinated or bot activity, confidence thresholds and near-real-time processing.

Value

Research teams can monitor far more conversation than they could read, spot emerging narratives early and focus human review where it matters.

The challenge

Front desks spend much of the day answering the same questions and relaying simple requests: breakfast times, fresh towels, late checkout, a broken air conditioner.

What we built

A messaging concierge guests use instead of calling reception. It answers property questions instantly and turns requests into tasks for the right department.

How a request flows

1. Guest message
2. Intent and context
3. Property knowledge
4. Task in service workflow
5. Staff action or escalation
6. Status update to guest

Key features

1. Instant answers on amenities, dining, policies and local recommendations.
2. Natural-language requests with no menus.
3. Automatic routing to housekeeping, maintenance or the front desk.
4. Multilingual conversations for international guests.
5. Human handoff with full context for unusual or sensitive situations.
6. Confirmations and status updates as requests progress.

Value

Fewer repetitive calls, faster service and more consistent coordination between teams.

The challenge

A leading Indian fashion platform adds thousands of products a week. Manual tagging of color, pattern, sleeve, fit and occasion is slow and inconsistent, which weakens search, filters and recommendations. Shoppers also struggle to describe a look in words.

Our role

Our team led machine learning and computer vision work at a large fashion e-commerce platform.

How it works

1. Attribute classification: models predict product attributes directly from catalog images.
2. Confidence-based review: low-confidence predictions go to a human queue; high-confidence ones publish automatically.
3. Image embeddings place every product in a visual similarity space.
4. Visual search: shoppers upload a photo and see visually similar products.
5. Similar-item recommendations on product pages, driven by the same embeddings.

Engineering considerations

Varied photography styles, long-tail attributes, cold start for new products and fast similarity search over a large catalog.

Value

Richer, more consistent catalog data, better filtering and search, and a new path to discovery through images.

The challenge

Medical knowledge is spread across millions of papers and clinical records written in dense, inconsistent language. Researchers cannot read it all, and simple keyword search misses synonyms, abbreviations and context.

Our role

Our team led NLP and data science work at a healthcare AI company.

How it works

1. Biomedical entity recognition for diseases, drugs, genes, symptoms and procedures.
2. Normalization of synonyms and abbreviations to standard vocabularies.
3. Relationship extraction to capture links such as drug to adverse event or gene to disease.
4. Context detection for negation, uncertainty and family history in clinical notes.
5. Knowledge graph and search so researchers can query relationships across the literature.

Engineering considerations

Domain-specific language models, privacy controls for clinical data, expert annotation and validation.

Value

Researchers find relevant evidence in minutes instead of weeks, with every result traceable to its source.

The challenge

Large enterprises spend heavily on technology, but the data arrives as messy general-ledger lines, invoices and cloud bills with inconsistent vendor names and descriptions. Mapping it to a clear cost model is slow, manual work.

Our role

Our team led machine learning work at an enterprise IT financial-management software company.

How it works

1. Text classification maps each transaction to a standard IT cost taxonomy.
2. Vendor normalization resolves name variants to a single vendor record.
3. Forecasting of spend by category and vendor.
4. Anomaly detection flags unexpected jumps in cloud or license costs.
5. Human feedback loop where analyst corrections retrain the models.

Engineering considerations

Short, noisy text, taxonomies that differ by customer, and explainable predictions finance teams can trust.

Value

Faster, more consistent cost allocation and earlier warning of overspend.

The challenge

High-volume communications platforms handle enormous amounts of voice and text in many languages, often mixing scripts and languages in a single message. Analysts need to find what matters quickly.

Our role

Our team led data science and NLP work at a large-scale communications analytics company.

How it works

1. Speech-to-text across multiple Indian and international languages.
2. Language and script detection, including code-mixed text such as Hinglish.
3. Entity extraction for names, places, organizations and numbers.
4. Topic and keyword discovery across conversations.
5. Scalable indexing and search so analysts can query large archives fast.

Engineering considerations

Noisy audio, transliteration, very high throughput and strict access controls.

Value

Analysts find relevant conversations in a fraction of the time manual review would take.

The challenge

Financial institutions process large volumes of forms, statements, contracts and correspondence. Manual data entry is slow, error-prone and hard to audit.

Our role

Our team led analytics and machine learning consulting work at a global professional-services firm.

How it works

1. Document classification sorts incoming files by type.
2. OCR and layout analysis read scanned and digital documents, including tables.
3. Field extraction captures names, amounts, dates and clauses into structured data.
4. Validation rules cross-check extracted values against source systems.
5. Exception routing sends uncertain cases to reviewers with the evidence highlighted.

Engineering considerations

Poor scan quality, many template variants, audit trails and data security.

Value

Faster processing, fewer errors and a clear record of every decision.