OjasSuratkar.

A high school transcript arrives as a scanned PDF and an admissions officer re-keys every course, grade and credit by hand, then it happens again at the next institution.

A platform that pulls the structured record out of the scan, verifies each requester with a one time code, and moves it through review across three roles.

• ExtractionLayout aware OCR into a typed record, not loose text.
• ValidationSchema check before persistence, so a bad parse fails loudly.
• InfrastructureCloud SQL and Cloud Run in Terraform, Alembic migrations.

The hard part is not reading a transcript, it is trusting it: extraction proposes, the schema decides, and only then does anything persist.

Pay gaps persist between people doing similar work, and setting pay from salary history carries an old gap into every job that follows.

A study over the NLSY97 job level panel, roughly 170,000 observations from 1997 to 2021, with model summaries and a policy counterfactual in a dashboard anyone can open.

• Panel24 years of job level records cleaned to consistent fields.
• EstimationJob and industry fixed effects: a 10.2 percent female penalty.
• CounterfactualCalifornia’s 2018 salary history ban as the experiment.
• DashboardChart.js, static, nothing to run to interrogate the result.

Headline gap figures compare across the labour market, which mixes the gap with how people are sorted into jobs. Holding job fixed answers the narrower and harder question.

A chat assistant summarises a document and then forgets it. The source is gone, and nothing stays searchable across everything you have read.

A workspace that keeps the originals. Documents, recordings, video and web pages stay playable and searchable, with summaries beside them and analysis across the whole collection.

• ProvenanceSmith-Waterman alignment with affine gaps, returning character offsets.
• ChronologyA grammar parses dates out of text and orders every mention.
• AudioSTFT fingerprinting locates a clip inside a long recording.
• IntegritySHA-256 hash chain with Merkle proofs, CRDTs for merges.

A chat assistant summarises better, and the project says so. What it cannot do is keep your material, or name the passage a claim came from.

Vendors, production, warehouses and distributors each keep their own spreadsheet, so an order at one end tells nobody at the other what to make or buy.

Four role dashboards over one schema. A distributor order releases warehouse stock, a restock becomes a production run, and the run derives its raw materials from the product recipe, with SKU level demand forecasting and what if simulation on top.

• DistributorOrders from the warehouse that covers its postcode.
• WarehouseAccepts against live stock and decrements in the same step.
• ProductionDerives material needs from the recipe and flags shortfalls.
• VendorFills the raw material requests that shortfall creates.

The dashboards are views onto one order trail rather than four applications, which is why a single order moves stock, opens a run and raises a purchase with nothing typed twice.

Riders cannot tell which bus serves their trip or how full it will be. The app that answered this had one city written into five screens as stop names, coordinates and routes.

A rider and conductor app that matches a bus to a start and end stop, follows it on a live map and predicts crowding by time slot. It now runs in any city.

• Line buildingClustering, then lines capped at a realistic stop count.
• Trip planningBreadth first across the network, fewest changes, either direction.
• RoutingOSRM through the backend, so no map key ships in the bundle.
• CrowdingHistorical ridership per slot, plus live density within 50 m of a stop.

Moving the city into a data layer is what makes it a transit app rather than a Mumbai app: point CITY at Berlin and the lines rebuild from OpenStreetMap.

Students take courses in one system and plan careers in another. Career advice stays generic and the skills on a resume are self reported.

A Canvas style LMS where every graded assignment is tagged with the skills it builds, so grades become an evidence backed profile matched against twelve career roles.

• Skill matrixGraded submissions become evidence, weighted by recency.
• CopilotBM25 fused with Gemini embeddings by reciprocal rank fusion, cited.
• IntegrityA request to solve open graded work switches it to coach mode.
• PlannerEarliest deadline first, with overload flagged from slack.

Readiness is computed from graded work rather than self report, which is the argument of the project. Every AI feature has a deterministic fallback, and responses land near 20 ms where the first version chained five model calls.

A claim spreads faster than the check on it, and the check usually arrives as a verdict with no working you can inspect.

An end to end NLP pipeline that classifies an article as fake or real, shipped twice over: a CLI predictor for batches and a Streamlit app for interactive use.

• PreprocessingCleaning, tokenising and feature extraction over raw article text.
• TrainingClassifier fitted and evaluated against held out data.
• AnalysisError patterns reviewed to guide the next retraining pass.
• DeliveryA CLI predictor and a Streamlit demo over the same model.

Shipping the model behind two interfaces rather than one is what made it usable: the CLI for volume, the app for the single article someone is actually arguing about.