Hi, I&#x27;m Kanishk Kapoor

Dual-model fallback

News Intelligence Dashboard

Real-time data pipeline: news API → dual-model summarisation (OpenAI GPT-4 + HuggingFace DistilBART fallback) → Streamlit dashboard with smart API rate-limit handling.

AI / LLM

PythonGPT-4HuggingFaceStreamlit+2

Multi-layer Medallion

Project Aeroflow — Real-Time Pipeline

End-to-end real-time airline delay data pipeline: FastAPI producer → Apache Kafka → Azure Event Hubs → Databricks PySpark (Bronze/Silver/Gold) → Snowflake → live Snowsight KPI dashboards.

Data Engineering

PythonKafkaAzureDatabricks+2

E-Commerce Sales Pipeline

Real-time order streaming pipeline: FastAPI event source → Kafka → Spark stream processing → structured JSON in AWS S3 with Airflow orchestration and Docker containerisation.

Data Engineering

KafkaSparkAWS S3Airflow+2

5M+ records

European Water Quality ML Model

Research-grade ML pipeline on 5M+ European environmental records. Spatio-temporal feature engineering, gradient boosting for nitrate/phosphate pollution risk prediction across 4 water body types.

Machine Learning

PythonXGBoostScikit-learnPandas+1

Transaction Fraud Detection

FinTech fraud detection on 18K+ transactions with feature engineering, statistical validation (ANOVA, Mann-Whitney U), XGBoost with deliberate class-imbalance handling.

Machine Learning

PythonXGBoostANOVAScikit-learn+1

View All Repos on GitHub

Research

Quantified models of where agents change the economics of development, design and construction across Ireland and the UK.

Real Estate & Development

Let the agent write code: TUM's result that the paradigm beats the model

A 1,027-task benchmark across 37 IFC models found that letting an agent explore a model by running code beat static query generation by 37 points. The weaker adaptive model beat the stronger static one.

+36.8 to 38.5pp

Adaptive over static, p < 0.001

1,027

Benchmark tasks, 37 models

55 to 57%

Strict accuracy, still not enough

10 min read

A number is not an estimate: the multi-agent study that produced cost distributions instead

Twelve cost engineers, three real buildings, and a measured comparison against deterministic BIM, manual probabilistic estimating and single-prompt GPT-4. The agents won, and the ablation says why.

12.5%

MAPE, against 18.7 to 27.4%

4.2 min

Per estimate, from 18.5 to 68

$0.15

API cost per run

10 min read

One photo, one updated model: what the first MCP progress-monitoring paper actually shows

A research team wired Claude to Revit and Blender through MCP and closed the loop from site photograph to updated 4D model with nobody in the middle. The result is an existence proof, not a product.

4 stages

Fully autonomous loop

91.4%

Completion computed from one image

1 wall

Total experimental scope

A 7B model that beat Claude and GPT-5.2 at reading building regulations

Fine-tuning plus reinforcement learning on a narrow regulatory task outperformed frontier models zero-shot. The argument for training rather than prompting, and for keeping regulation on your own hardware.

−23.8%

Tree edit distance vs SFT baseline

−38.6%

Token-level Levenshtein distance

Parameters, locally hostable

The front half of compliance checking is nearly solved. The back half is where the work is

97% F1 on rule classification and 100% on dependency identification. Interpreting regulations turns out to be the easy part, and the industry has been optimising the wrong end.

97%

F1, rule classification

100%

Dependency identification

97%

Correct tool selection

Real Estate & Development

Five papers, one pattern: nobody won by using a bigger model

Reading a month of agentic AI research in the built environment side by side. The architectures differ completely and the conclusion is the same one every time.

5 papers

Read side by side

Won by scaling the model

Where data quality set the ceiling

7 min read

Site selection in eleven days: running development due diligence with agents

A worked model of agentic site screening for Irish and UK development, including where the time actually goes and which parts refuse to automate.

14 wks → 11 days

Brief to ranked shortlist

Sites screened per cycle

€31k

Modelled saving per scheme

Every facade option, priced four ways: predictive design past generative design

Generative design produced options nobody could evaluate. Predictive design forecasts how each option performs structurally, financially and in embodied carbon before anyone commits.

4 min

Per option, fully appraised

217 kgCO₂e/m²

Spread across options

6.2%

Capital cost variance found

10 min read

The first draft of every RFI: document agents on live construction projects

Site teams lose more hours to structured paperwork than to any technical problem. That work has a shape, and the shape is automatable.

11.4 hrs

Weekly admin recovered per PM

2.1 days

RFI turnaround, from 6.8

412 pp

Tender reviewed in one pass

Real Estate & Development

The lease portfolio that reads itself: compliance agents across commercial estates

Rent review triggers, break clauses and statutory deadlines are a data extraction problem that portfolios keep solving with calendar reminders and hope.

1,240

Leases parsed per run

94%

Clause extraction accuracy

£410k

Missed escalation exposure found

Agents that drive the CAD software: what Model Context Protocol changes for design practices

The interesting shift is not agents that describe engineering work. It is agents that operate the specialist tools directly, and the governance that has to come with it.

71%

Modelling tasks agent-operable

3.4×

Option throughput at stage 3

100%

Actions requiring audit trail

Closing the loop between site and model: reality capture as a live data asset

Drone and 360 capture produces enormous datasets that mostly get archived. Agents turn that data into deviation detection, and deviation detection into design action.

8 days → 4 hrs

Capture to deviation report

23mm

Detectable deviation threshold

£1.9m

Modelled rework avoided