Now
AI Engineer, Tazapay
AI for fraud detection and anti-money laundering in cross-border payments.
p.01 · Notebook Nº 03 · Bangalore · Tonight's set
Samarth Saraswat,
AI builder.
SetupI’ve got the Heart of a Product Builder,Brain of an Engineerand Greed of a Salesperson
Punchbecause passion doesn't pay the rent.
What can I do for you?
Bring me the messy problem. I'll build the AI agent, the RAG, the model, or talk you out of all three.
TagIIT Guwahati taught me to take problems seriously. Stand-up Comedy taught me not to take myself seriously.
The day job, briefly
Before · 3+ years
Data Scientist, Microsoft
Built Seller Copilot, an AI for enterprise sales. Stopped $100M+ a year in revenue leakage.
Go to my LinkedIn if you care about my job.
p.02
p.02 · Who
The fine print.
- Now
- AI Engineer, Tazapay
- Before
- Data Scientist, Microsoft · 3+ years
- School
- B.Tech, Electronics & Communication, IIT Guwahati, 2022
- Also
- YC Startup School, Bangalore · the first edition there, April 2026
- Based in
- Bangalore, India
- Writes at
- samarthsaraswat.substack.com
- Off hours
- Stand-up comedy. The mic is not a prop.
Index card · the facts
Buzzwords I can back up
AI systems
LLM copilots & agentsRAG & hybrid retrievalOn-device inferenceEvals & golden setsFraud & AML detection
Machine learning
NLPSpeech recognitionAnomaly detectionModel evaluationDeep learning
Product
AI product managementExperimentationOKRs & North Star metricsMetric design
Tools
PythonPyTorchLangGraphClaude CodeCodexSQL
p.03
p.03 · Built · Open source
Things I built, then gave away.
Two of these run entirely on the device in your hand or on your desk, with nothing sent anywhere. The third is a free study hub. All three are on GitHub.
Exp.01 · On-device AI · Android
Survive AI
An offline-first survival assistant that runs a 2B language model and a retrieval pipeline entirely on an Android phone.
fig. 2 · On-device AI · Android
works with zero bars!
Lab results
- Runs on
- Android, 6 GB RAM, zero bytes sent at runtime
- Model
- Gemma 2B IT, INT4, CPU backend
- Retrieval
- Recall@5 89.7% hybrid vs 81.5% keyword-only
- Stack
- Flutter, SQLite FTS5, ONNX Runtime, Python build step
+read the full lab notes
Emergencies take the network with them. Floods drop cell towers, shutdowns remove the rest, and the moment someone needs to know what to do about a snakebite or a gas leak is the moment they cannot look it up.
Survive AI answers “kutte ne kaata” or “chest pain” with no bars. Eighteen India-specific guides are searched with three retrieval legs (BM25, query expansion with romanised Hindi, and a dense encoder), fused with weighted reciprocal rank fusion, and handed to Gemma 2B through an instruction-last prompt tuned for small models. Every answer cites the paragraph it came from, and an output guard blocks anything the guides forbid.
Exp.02 · Local speech · macOS
Desi Dictation
Local-first Hinglish dictation for macOS. Hold a key, speak the way you actually talk, release, and Roman-script Hinglish lands in whatever app you are typing in.
fig. 3 · Local speech · macOS
1.47 s on an M3 Air
Lab results
- Runs on
- macOS, Apple Silicon, fully local
- Latency
- 0.21 s English, 1.47 s Hinglish per dictation (M3 Air)
- Models
- Oriserve Whisper-Hindi2Hinglish, Parakeet TDT, Vaani
- Stack
- Swift, SwiftUI, whisper.cpp, Metal, Silero VAD
+read the full lab notes
Dictation tools understand English or Hindi. Nobody in Bangalore talks like that. “Kal meeting hai, please deck ready rakhna” is one sentence, and it should land as one sentence.
A global hotkey captures 16 kHz audio, whisper.cpp runs a Hinglish fine-tuned Whisper on Metal, a user dictionary cleans it up, and a pasteboard swap drops the text into the focused app. No cloud, no subscription, runs on a MacBook Air. The current beta ships five curated models with per-language defaults, voice-activity pause handling, chunked transcription for long dictation, and a warm-mic pre-roll so the first word is never lost.
Exp.03 · Community · Free study hub
Data Science Preparation
A free, hosted study hub for anyone preparing for data science and machine learning roles.
fig. 4 · Community · Free study hub
free. no sign-up. ever.
Lab results
- Lives at
- GitHub Pages, rebuilt on every push
- Format
- 7 MkDocs Material sites, one hub
- For
- Anyone. No sign-up, no paywall
+read the full lab notes
Seven self-contained knowledge bases, built as independent MkDocs sites and stitched under one landing page: Mathematics, Classical ML, Deep Learning, LLMs, Agentic AI, Productionizing ML, and twenty ML system-design case studies with full interview transcripts.
It also doubles as a knowledge base for AI agents. A machine-readable index maps every topic to its page, with a citation protocol so an agent can say what it already knew versus what it went and looked up.
p.04
p.04 · Writes · Substack
Things I wrote down.
Product thinking, AI strategy, and metric design.
p.05
p.05 · Crowd work · Raise a ticket
Got a messy problem? Raise a ticket.
AI product work, collaborations, or a friendly argument about where AI is headed. The ticket lands straight in my inbox.
Or the old-fashioned way
- linkedin.com/in/s-samarth
- GitHub
- github.com/s-samarth
- Phone
- +91 70623 03003
Support ticket
Nº ····