EXP-014SPEECH · PEFT
Can a 1B-parameter speech model be adapted using only a tiny fraction of its parameters?
Uyghur ASR, one Tesla T4, ~23h of audio.
- RESULT
- 0.005% parameters trained
- COMPUTE
- 1 × Tesla T4 · ~1 hour
- OUTCOME
- CER = 0.0517
AVAILABLE FOR ML / AI OPPORTUNITIES
I build machines
that learn.
Machine Learning Engineer building AI systems across language, vision, speech, and data.
“From fine-tuning speech models on a single GPU to building document intelligence systems, I like working at the intersection of research, mathematics, and engineering.”
computational_graph.pyLIVE ●
INPUTMOVE THE CURSOROUTPUT
shramadeep@lab: ~
❯▌
·▌
·▌
❯▌
·▌
·▌
·▌
·▌
·▌
❯▌
·▌
·▌
·▌
·▌
INTERACTIVE · CLICK REPLAY TO WATCH AGAIN
ROLEML Engineer · Researcher · Data Scientist
DOMAINSNLP · CV · Speech · LLM Systems · Data Science
STATUSBuilding, measuring & shipping
Building intelligent systems from models to production. — every project below is a problem I chose, measured, and learned from.
quick stats
00+
Years building ML systems
Since the first model went from notebook to code
00+
Research & engineering projects
Across language, vision, speech, and data
00+
AI domains explored
NLP · CV · Speech · Time-series
00
GPU used for recent ASR fine-tuning
Tesla T4 · ~1 hour of training
01about
A little about how I work
I'm a Data Science student, interested in understanding how intelligent systems actually work — from the mathematics behind statistical learning to the engineering required to deploy models reliably.
My work spans NLP, computer vision, speech recognition, LLM applications, data analytics, and machine learning infrastructure.
I particularly enjoy problems where the obvious solution isn't enough: long-context document processing, low-resource speech recognition, OCR pipelines, model fine-tuning under limited compute, and systems that have to work outside a notebook.
area of focusNODE MAP
NLP & LLMslanguageComputer VisionvisionSpeech AIspeechDocument IntelligencesystemsData SciencedataML Infrastructuresystems
HOW I CHOOSE WORK
- 01Problems the obvious solution can't solve
- 02Stories that survive measurement
- 03Systems that leave the notebook
02how i think
I don't just use models. I try to understand them.
Each stage conditions the next. I go down this chain on purpose — when a model misbehaves, I go back up it to figure out which layer lied.
STAGE 01 / 07
Mathematics
The substrate everything else sits on.
01Linear algebra
02Probability
03Optimization
04Calculus
03selected work
Selected Work
Things I've built, trained, broken, optimized, and learned from. Each card expands into a full case study — what the problem was, how I approached it, what happened, and what I'd do differently.
04experiments
Experiments
A running research notebook — small, self-contained probes with a question, a result, and a number when I have one. Not everything works; that's the point. (unverified cells read [ADD …])
EXP-009EFFICIENCY
Where does fine-tuning actually spend its trainable parameters — and what does LoRA keep?
LoRA vs full fine-tune on the same task and data budget.
- RESULT
- ~100× fewer trainable params
- COMPUTE
- 1 × consumer GPU
- OUTCOME
- Within noise of full fine-tuning [ADD METRIC]
EXP-008ASR · THEORY
How does CTC solve frame-to-text alignment without attention?
Reading along with the model internals, reproducing alignment behaviour.
- RESULT
- Blank token absorbs excess frames
- COMPUTE
- —
- OUTCOME
- A working mental model of CTC decoding
EXP-005VISION
Are CNNs really that much more parameter-efficient than MLPs per unit of computation?
Small-scale comparison: receptive fields vs dense connectivity.
- RESULT
- O(k²·c) vs O(n²) weight growth
- COMPUTE
- —
- OUTCOME
- Quantified why spatial structure wins
EXP-004DATA
How much of the variance actually survives a PCA projection?
High-dimensional tabular data, k components chosen by explained variance.
- RESULT
- [ADD VALUE]% variance retained
- COMPUTE
- —
- OUTCOME
- k = [ADD VALUE] components → [ADD VALUE]% drift-free reconstruction
EXP-002FORECASTING
Can ARIMA and Prophet actually beat a naive seasonal baseline on retail demand?
SKU-level sales history; strict backtesting.
- RESULT
- Both do, on the right SKUs [ADD METRIC]
- COMPUTE
- —
- OUTCOME
- Map of which series deserve deep models
EXP-001SYSTEMS
What actually changes between fp16 and int8 for a served transformer?
Size, latency, and output-quality deltas on a real endpoint.
- RESULT
- [ADD %] smaller · [ADD ×] faster
- COMPUTE
- local + serverless
- OUTCOME
- Quantise only where quality holds
EXP-011RETRIEVAL
At what chunk size does retrieval quality start to degrade?
Same corpus, embeddings, and query set — swaplling chunking strategy only.
- RESULT
- Layout-aware chunking > fixed windows [ADD METRIC]
- COMPUTE
- vector DB on free tier
- OUTCOME
- Chunking rules adopted across projects
// experiments are logged the way I run them — one question, one controlled variable, one honest number.
05engineering
Beyond the model
Training a model is a small fraction of making it useful. The rest is the unglamorous lifecycle — data, latency, servers, and monitoring.
ml lifecyclePIPELINE
01
Datacollection · cleaning · labelling
02
Processingpipelines · ETL · augmentation
03
Modeltraining · fine-tuning · evaluation
04
APIserving · FastAPI · REST
05
DatabasePostgreSQL · MongoDB · vector DBs
06
DeploymentDocker · Linux · cloud
07
Monitoringdrift · latency · logging
tools I reach for
FastAPIFlaskDockerLinuxMongoDBPostgreSQLSQLiteCloudflare tunnelsCoolifyVercelAWS
Most ML work isn't gradient updates — it's the unglamorous plumbing around the model. I care about the whole lifecycle because a model that can't be served, evaluated, or monitored is just a research artefact.
I build REST APIs with FastAPI, containerize with Docker, deploy on Linux boxes and apps like Coolify or Vercel, and wire vector stores for retrieval. The goal: ideas that survive contact with the real world.
PRINCIPLES
- A model you can't serve is a research artefact.
- Evaluate before you optimise — and after.
- The pipeline is a product; the model is a component.
06technical stack
Tools I reach for, as a map
Everything orbits a core: understanding models well enough to build systems around them. Hover a node to see its orbit.
stack mapHOVER A NODE
Languages
PythonJavaScriptTypeScriptSQL
ML / AI
PyTorchTensorFlowScikit-learnTransformersPEFTLoRA
NLP
LLMsRAGOCREmbeddingsVector Search
Computer Vision
YOLOCNNsSwinIRImage Processing
Data
PandasDuckDBSQLTime SeriesStatistics
Backend
FastAPIFlaskREST APIs
Frontend
ReactNext.jsVue
Infrastructure
DockerLinuxCloudflareCoolifyGit
HUBML Engineering
ML Engineering
Languages
PythonJavaScriptTypeScriptSQL
ML / AI
PyTorchTensorFlowScikit-learnTransformersPEFTLoRA
NLP
LLMsRAGOCREmbeddingsVector Search
Computer Vision
YOLOCNNsSwinIRImage Processing
Data
PandasDuckDBSQLTime SeriesStatistics
Backend
FastAPIFlaskREST APIs
Frontend
ReactNext.jsVue
Infrastructure
DockerLinuxCloudflareCoolifyGit
07knowledge graph
My technical map
Hover a concept — its branch and children light up. This is how I think about the field: a graph, not a list.
technical_map · interactiveHOVER TO TRACE CONNECTIONS
CORESPECIALISATIONS
08experience
Where I've built things
A short, honest timeline — not a resume dump. Expand for what I actually did.
Working on production-oriented ML systems — applying models to real product problems end to end.
- ›[ADD RESPONSIBILITY] · model work, data, or evaluation you owned
- ›[ADD RESPONSIBILITY] · measurable outcome e.g. accuracy / latency / cost
- ›[ADD RESPONSIBILITY] · engineering you shipped (APIs, pipelines, deployment)
09education
Theoretical grounding
Indian Institute of Technology Madras
BS in Data Science · B.Sc. Data Science & Applications
focus areas
StatisticsMathematicsMachine LearningData ScienceComputer ScienceEconomics / Finance
why it matters here
Every model is a statistical claim. The program builds the math under the machine learning — so when a network misbehaves, I can interrogate it at the level of loss surfaces and priors instead of guessing hyperparameters.
representative coursework
01Probability & Statistics✓
02Linear Algebra✓
03Machine Learning Foundations✓
04Data Structures & Algorithms✓
05Database Systems✓
06Intro to Deep Learning✓
07Time Series & Forecasting✓
08Business Finance✓
10currently exploring
Currently exploring
Areas I'm actively reading, experimenting, and building in — not claims of expertise. Curiosity is a work-in-progress by definition.
01
+Reinforcement Learning
reward design · policy learning
02
+Multimodal AI
language × vision × audio
03
+Agentic Systems
tool use · planning loops
04
+Efficient Fine-Tuning
PEFT · quantisation
05
+AI Security
prompt injection · evals
06
+HFT / Quantitative ML
signal extraction at scale
11notes from the lab
Notes from the lab
Technical notes written to actually understand things — short explanations of concepts I've used in projects. Published from plain markdown.
12open source
Open source & code
Repositories mirror the projects on this page. Published code, honest descriptions — stars and dates only appear when they're real.
the condensed version
Want the condensed version?
One page, the way recruiters want it — experience, projects, education.
let's connect
Let's build
something interesting.
Interested in ML systems, research, AI engineering, or difficult technical problems?
I reply to real messages about models, systems, and hard problems — usually within a day.