AI engineer · MSc Artificial Intelligence candidate at NTU Singapore
I like AI most when it crosses disciplines: process physics becomes a fast design tool, a wafer image becomes an inspectable review signal, or a multilingual recording becomes useful without leaving the machine. I build these systems from mechanism to evidence to interface, then document where they stop being reliable.
Email · LinkedIn · Start a technical conversation
Semiconductor intelligence · Speech and language · Vision and multimodal AI · Evaluation and trust
Semiconductor intelligence
Physics-guided learning, wafer-map understanding, surrogate modelling, uncertainty, and engineer-facing evidence.
Speech and language
Local multilingual transcription, translation, diarization, retrieval, and privacy-conscious pipelines.
Vision and multimodal AI
Computer vision, synthetic-media analysis, 3D reconstruction, diffusion models, and evidence fusion.
Evaluation and trust
Grouped splits, calibration, abstention, distribution-shift awareness, provenance, and claims that match the evidence.
This section refreshes automatically from my latest substantial public repositories. A new project with a clear GitHub description joins the feed after its first push.
Harshsaand.Github.Io
Harsh Saand's public AI portfolio.
HTML / Aug 2026
Processtwin Semiconductor Ai
Physics-grounded surrogate modelling for silicon oxidation, dopant diffusion, uncertainty and inverse process design.
Python / inverse design / physics informed ml / Aug 2026
Lithotwin Computational Lithography
Conditional neural surrogate for computational lithography and resist-contour prediction.
Python / Aug 2026
Wafer Process Signature Triage
Interpretable wafer-map classification with spatial signatures, calibrated confidence and similar-case retrieval.
Python / anomaly triage / computer vision / Aug 2026
Local Multilingual Speech Intelligence
Privacy-conscious local multilingual transcription, translation and optional speaker diarization with structured exports.
Python / local ai / multilingual / Aug 2026
Multi-Agent Tileworld
Cooperative Tileworld agents using A*, shared working memory, broadcasts, complementary patrols and fuel-aware planning.
Java / a star / cooperative ai / Aug 2026
These are not the limits of what I build. They are the clearest examples of how I think.
ProcessTwin AI TCAD | physics-guided semiconductor modelling
Can reduced-order silicon process physics and a learned surrogate support fast recipe exploration without hiding what the model approximates?
recipe → reduced-order physics → simulated DOE → PCA + residual ensemble → inverse search → solver verification
- Built Deal-Grove oxidation, 1D dopant diffusion, a deterministic 6,000-recipe simulated DOE, PCA profile compression, residual MLP ensembles, OOD warnings, and simulator-verified inverse design.
- Held-out simulator fidelity reached R² 0.9968 for oxide thickness and 0.9985 for junction depth. These are simulator-backed results, not fab calibration.
- Inspect the source, saved metrics, tests, or the technical report.
Wafer Process Signature Triage | spatial evidence for engineering review
Can wafer-map geometry become a useful review signal while keeping spatial structure visible to an engineer?
wafer image → grouped split → Cartesian and polar views → calibrated ranking → review route + similar cases
- Fused Cartesian and polar CNN views with a 17-value spatial signature, grouped splitting, temperature calibration, uncertainty routing, and similar-case retrieval.
- The saved seed-42 run reached 0.812 macro-F1 on an untouched 138-image grouped test split across nine classes.
- This is pattern triage, not causal root-cause diagnosis. Inspect provenance, source, metrics, or tests.
DeepShield ApprovalGuard | multimodal integrity evidence
Before a sensitive financial instruction moves forward, can local AI surface media-integrity evidence that deserves human review?
recording → quality checks → independent evidence branches → timestamped timeline → review route
- Combined synthetic-voice, face-manipulation, visual-continuity, media-quality, and audio-video timing branches with timestamped evidence and abstention.
- The AASIST audio branch reached ROC-AUC 0.9078 on a balanced 570-file ASVspoof subset. Other branches use functional fixtures and are not presented as validated detectors.
- The system supports review; it does not establish identity, intent, or fraud. Inspect the pipeline, evaluation, dataset card, or model provenance.
Local Multilingual Speech Intelligence | private speech pipelines
How can multilingual recordings be transcribed, translated, speaker-labelled, and reviewed while keeping audio processing local?
local media → VAD + faster-whisper → optional diarization → overlap assignment → TXT / JSON / HTML
- Built a compact
faster-whisperpipeline with optionalpyannote.audiodiarization, temporal-overlap speaker assignment, and accessible exports. - The repository includes synthetic data and deterministic tests for timestamps, speaker overlap, JSON preservation, and HTML escaping. It is a reference pipeline, not a claimed speech-quality benchmark.
- Inspect the source, synthetic output, or tests.
find the failure mode
↓
build the smallest measurable system
↓
inspect errors and distribution shifts
↓
state what the result does not prove
↓
design the path to use
Mechanism before mystique. I want to know what produces an output and where the abstraction breaks.
Evaluation before adjectives. A metric needs a dataset, protocol, baseline, and honest boundary around it.
Privacy is architecture. Data handling and deployment constraints shape a system from the beginning.
Interfaces are part of the model. Evidence matters when another person can inspect it and make a better decision.
More experiments and earlier builds
My broader work includes multi-agent Tileworld, retrieval-augmented generation, diffusion style transfer, disaster-tweet classification, and geometric 3D reconstruction. Some are compact experiments; others grew into the systems above.
- How can mechanistic models and learned surrogates work together with uncertainty that is actually calibrated?
- How should multimodal integrity systems behave under distribution shift, missing evidence, and deliberate attack?
- How can multilingual speech systems remain private, inspectable, and useful on constrained hardware?
Learning systems PyTorch, Transformers, RAG, LoRA/PEFT, model evaluation
Speech and NLP faster-whisper, WhisperX, pyannote, multilingual translation
Vision CNNs, diffusion models, geometric vision, 3D reconstruction
Engineering Python, Java, REST APIs, Docker, Kubernetes, GPU computing
I am open to internships and early-career roles in applied AI/ML and research engineering, especially in semiconductor intelligence, speech and NLP, computer vision, multimodal systems, and trustworthy evaluation.
If our questions overlap, email me, connect on LinkedIn, or start a technical conversation.