TTwissa Modi
Open to AI / ML roles
Let's talk
Twissa Modi · Applied AI / ML engineer

AI systems that hold up in the real world.

I build forecasting, vision, and AI agent systems. Previously at Microsoft; now working on AI agents at Bosch while studying AI at NTU.

See selected work More about me
Based in SingaporeM.S. AI at NTUModels → services → interfaces
01 / proof16%compute-cost reduction on a Microsoft backend system
02 / proof2.65%test MAPE on a day-ahead electricity forecast
03 / proof84%test accuracy on a deployed wildlife classifier
04 / proof3+ yrsshipping production software before specializing in AI
Selected work / 01

Proof over promises.

A few systems across forecasting, computer vision, and agentic workflows. Each project documents the constraint, the decisions, and what actually shipped.

View all projects
01 / Deep Learning / TransformersPublished

Decoder-Only Transformer: MHA vs GQA vs MQA

A GPT-style decoder-only Transformer built from scratch in PyTorch to study the trade-off between language-model quality and autoregressive inference efficiency.

PythonPyTorchTransformersAttentionRoPEDeep LearningCUDA
02 / Forecasting / ReliabilityPublished

Day-ahead electricity load forecasting

A probabilistic forecasting system designed around the realities of a day-ahead horizon, not just a leaderboard score.

Primary result
2.65%test MAPE
Time seriesCQRMLOps
03 / Computer vision / ProductPublished

A wildlife classifier that shows its work

A full data-to-demo computer vision system with honest splits, explainable predictions, and a usable interface.

Primary result
84.0%test accuracy
Computer visionGradCAMFull stack

More work is being documented as it becomes useful to share, complete with the artifact, the tradeoffs, and the limits.

Open the project archive
The throughline / 02

Production depth. Applied AI direction.

The strongest work happens where model quality, system reliability, and user judgment are treated as one problem.

At Microsoft, I built product telemetry and backend services. At Bosch, I’m working on natural-language data analysis and agent interoperability alongside my M.S. in AI at NTU.

01 / EngineeringServices that stay dependableBoundaries, cost, observability, deployment
02 / ModelingEvaluation that respects realityHonest splits, calibration, known limits
03 / CommunicationInterfaces that earn trustClear writing, useful surfaces, explainable output
Read the longer version
Writing / 03

Making technical ideas easier to see.

Essays and explanations for the concepts that deserve more than a formula and a footnote.

Browse writing
article2026

Why Initialization Matters in a Transformer

When my from-scratch Transformer started training with an unexpectedly high loss, I traced the problem back to parameter initialization. This is what I learned about initialization, residual connections, and why GPT-style models use depth-aware scaling.

Read article
article2026

L1 vs L2, the intuition most explanations miss

Geometry, normalization, and the practical tradeoffs behind L1 and L2 distances.

Have a problem worth exploring?

Let's make it useful.

I'm looking for AI / ML engineering roles where I can develop models, evaluate them, and build the services around them.

Start a conversation